Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Models for predicting treatment efficacy of antiepileptic drugs and prognosis of treatment withdrawal in epilepsy patients

Models for predicting treatment efficacy of antiepileptic drugs and prognosis of treatment... Although antiepileptic drugs (AEDs) are the most effective treatment for epilepsy, 30–40% of patients with epilepsy would develop drug-refractory epilepsy. An accurate, preliminary prediction of the efficacy of AEDs has great clinical significance for patient treatment and prognosis. Some studies have developed statistical models and machine- learning algorithms (MLAs) to predict the efficacy of AEDs treatment and the progression of disease after treatment withdrawal, in order to provide assistance for making clinical decisions in the aim of precise, personalized treatment. The field of prediction models with statistical models and MLAs is attracting growing interest and is developing rapidly. What’s more, more and more studies focus on the external validation of the existing model. In this review, we will give a brief overview of recent developments in this discipline. Keywords: Prediction model, Machine learning, Antiepileptic drugs, Drug response, Withdrawal reaction Background burden on their families. Accurate prediction of the effi- Epilepsy is one of the most common neurological dis- cacy of AEDs before the initiation of treatment can reduce eases and has affected more than 68 million people the use of ineffective drugs, alleviate patients’ pain, and worldwide [1]. Although antiepileptic drugs (AEDs) are improve the prognosis in patients. currently the primary treatment option for patients with Doctors and patients often find it hard to decide epilepsy (PWE), about 40% of PWEs will suffer the con- whether to reduce or stop AEDs use. Although PWEs sequence of drug-resistant epilepsy (DRE) [2]. According can have better cognitive performances and higher qual- to the International League Against Epilepsy, DRE is de- ity of life after AEDs withdrawal, there is also an in- fined as the failure to achieve sustained seizure freedom creased risk of recurrence. To avoid recurrence, many after adequate trials of two tolerated and appropriately patients decide to put up with the side effects rather chosen AEDs treatments (monotherapies or combination than completely withdrawing the AEDs. Therefore, there therapies) [3]. The mechanism of DRE is not fully clear is an urgent need of effective and practical decision- and may be related to the sensitivity of drug targets, activ- making tools to assist clinicians to establish the course ity of drug transporters, cytochrome P450, structural of AEDs treatment and withdrawal, as well as to help neural network, and other potential causes of epilepsy [4]. realize precise, personalized treatment. Abrupt and repetitive seizures may lead to neurobiochem- Prediction models can integrate multiple clinical or ical changes in the brain, cognitive decline, and serious non-clinical parameters within a certain time to calcu- psychological problems in PWEs, which can seriously late the probability of diagnostic outcomes as well as the affect patients’ quality of life and cause an increased disease prognosis. These models can stratify patient risk stratification to support clinical decision-making and im- prove the prognosis and quality of care for patients [5]. * Correspondence: hanxiong7589@126.com Prediction models are divided into two main categories: Department of Neurology, People’s Hospital of Zhengzhou University, Henan Provincial People’s Hospital, Zhengzhou 450003, China © The Author(s). 2021 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Yang et al. Acta Epileptologica (2021) 3:1 Page 2 of 6 those based on statistics and those based on machine comorbidities, history of central nervous system infec- learning algorithms (MLAs). tion, cognitive impairment, epilepsy syndrome, presence The statistical prediction models are those whose de- of structural abnormalities in magnetic resonance im- velopment is based on statistics, such as univariate and aging (MRI), previous history of status epilepticus, family multivariate logistic regression/COX regression analysis history of epilepsy, history of perinatal brain injury, and to select prediction variables. The integration of multiple certain electroencephalography (EEG) features [12–15]. selection variables is used to calculate the probability of The identification and determination of these parame- a particular diagnosis or disease prognosis [5]. The de- ters are the basis for creating a predictive model. velopment of statistical prediction models involves col- lection of datasets, selection of prediction variables, Statistical prediction models development of a prediction model, evaluation of the Based on the statistical method, certain variables can be model’s performance, internal and external validation, selected and integrated into a model and a score system and further update of the model. This type of method can be created for a specific purpose; this type of model can be used to create an easy-to-use prediction scoring has been tested in the field of epilepsy. Boonluksiri et al. system [6]. One example is the Framingham risk score, [16] enrolled 308 children with epilepsy in a retrospect- which is widely used in the public health field for esti- ive study, and they selected the age at onset, prior mating the probability of the occurrence of cardiovascu- neurological deficits, and abnormal EEGs as variables, lar diseases in an individual within the next 10 years. and established a scale for predicting DRE in children. This model was built based on the traditional prediction The children were then divided into 3 groups depending variables of age, sex, systolic blood pressure, hyperten- on the risk of developing DRE: low risk (score < 6 sion treatment, total and high-density lipoprotein chol- points), moderate risk (score 6–12 points) and high risk esterol levels, smoking, and diabetes [7]. (score > 12 points), with positive likelihood ratios of 0.5, With the development of artificial intelligence, ma- 1.8, and 12.5, respectively, and an area under the curve chine learning is concerned with algorithm induction to (AUC) of 0.76. However, as this retrospective study was improve model performance by using statistical and conducted in a single center with a small sample size computer science approaches, and has shown potentials and lacked both internal and external validation, the per- for industrialization. Machine learning has also been ap- formance and practicality of this model need further val- plied to fields like speech recognition, image classifica- idation. In a previous study [15], we developed a scale tion, text translation, and medical care [8] for the for predicting DRE in adult patients with MRI-negative detection of critical findings in head computerized tom- epilepsy [MRI(−)DRE]. The AUC was 0.89 and the risk ography scans [9] and the classification of cancer [10]. stratification was given as: low risk (0–3 points), medium The machine learning technique is superior to manual as- risk (3–5 points), and high risk (> 5 points). Using this sessment by clinical experts in that it has higher accuracy method, the probability of DRE could also be calculated. of diagnosis and outcome prediction, and it can also be However, this scale was limited by the retrospective de- used for epilepsy, especially for automated seizure detec- sign based on data of 132 patients, so further validation tion, analysis of imaging and clinical data, epilepsy is needed. Latzer et al. [17] created a model to predict localization, and prediction of medical and surgical out- DRE in children with cerebral palsy at the Tel Aviv comes [11]. Additional validation techniques, such as the Medical Center in Tel Aviv, Israel and this model was hold-out cross-validation, k-fold cross validation, and used in a retrospective study including 118 patients. The “leave-one-out method” cross-validation, can be used to es- model was composed of four parameters (low Apgar timate the performance of the technique. In the following, score at 5 min, neonatal seizures, focal-onset epilepsy, we will give a brief overview of the efficacy of statistical pre- and focal slowing on EEG) and the AUC was 0.84. Al- diction models and MLAs for predicting AEDs treatment though their model helped to identify which patient response and patients’ outcome after AEDs withdrawal. would achieve better seizure control, the study was per- formed with a small sample size and the lack of valid- Prediction models for the response to AEDs ation makes it difficult to judge the model’s performance treatment (Table 1). Drug selection mainly relies on official guidelines and In summary, a few comprehensive statistical models clinical experience of doctors, due to the fact that the with multiple variables have been established for predict- treatment efficacy varies among individual PWEs. Per- ing the response to AEDs treatment, but they were sonalized selection of effective AEDs still remains a big weakened by some limitations including the retrospect- challenge. Some prospective studies have identified cer- ive design, small sample sizes, and the lack of internal tain predictive factors of DRE, such as early onset, sex, and external validation. To address these, more pro- duration of epilepsy, multiple seizure types, spective, multi-center studies with large sample sizes are Yang et al. Acta Epileptologica (2021) 3:1 Page 3 of 6 Table 1 Statistical prediction models for the response to AEDs treatment Study Study Prediction Development Final factors Validation AUC Stratified risks design target cohort cohort Boonluksiri Retrospective DRE in children 308 cases in Hatyai Age onset, prior neurological No 0.76 low risk: < 6 points; et al. 2015 [16] Hospital in Thailand deficits and abnormal EEG moderate risk: 6–12 points; high risk: > 12 points Yang et al. Retrospective DRE in adults 132 cases at Henan EEG before AEDs, history of No 0.89 low risk: < 3 points; 2019 [15] Provincial People’s CNS infection, initial precipitating moderate risk: 3–5 Hospital in China injuries, and more than one points; high risk: > 5 recurrence in the first 6 months points Latzer et al. Retrospective DRE in children 281 children with Low Apgar score at 5 min, No 0.68 No 2019 [17] with cerebral cerebral palsy at the neonatal seizures, focal-onset palsy Dana-Dwek Children’s epilepsy and focal slowing on EEG Hospital DRE Drug resistant epilepsy, AUC Area under the curve, EEG Electroencephalogram, MRI Magnetic resonance imaging, CNS Central nervous system required. In addition, these models need to be verified in random forest algorithm, to identify patients at high risk multiple centers. of DRE. A total of 292, 892 patients met the inclusion criteria for epilepsy; 175, 735 of them were assigned to Machine learning algorithms the training cohort and the other 117, 157 were assigned MLAs can be used to extract more EEG, imaging, and to the test cohort, and 1 270 features were screened as clinical features of patients to build prediction models predictive factors. The random forest algorithm had an and validate performance through the use of more AUC of 0.76 and performed the best of the three methods. MLAs can also be readily applied in artificial models. It could predict the emergence of DRE approxi- intelligence-based industrialization, an extremely rele- mately 2 years in advance before a patient failed two vant and competitive field today. AEDs trials. The drawback of this study was that it was UCB Pharma has been actively involved in conducting a retrospective study without external validation. Fur- research on the development and validation of MLA for thermore, the DRE incidence was only 13.1%, which was use in the prediction of AEDs effectiveness in individual lower than that in other studies, indicating that the data- PWEs. Devinsky et al. [18] at the New York University set had significant limitations. Medical Center, based on the UCB–IBM collaboration, A number of pharmacogenomic studies have focused explored the application of MLA to construct an algo- on identifying single nucleotide polymorphism (SNP) rithm for AEDs prescription. A total of 50, 000 PWEs markers for predicting the outcomes of AEDs treat- were retrospectively enrolled in the study and randomly ments, and some studies have tried to establish certain divided into a training group of 40, 000 patients and a multi-SNP models to predict the response to AEDs. Pet- testing group of 10, 000 patients. Roughly 5, 000 features rovski et al. [20] prospectively collected the genetic re- were extracted to build the prediction model, which had sults of patients with newly diagnosed epilepsy and an AUC of 0.72 and was considered to have a good pre- developed a multi-SNP classification model, based on dictive power. The patients with the model-predicted the k-nearest neighbor supervised learning approach, to AEDs regimen had significantly higher survival rates predict the seizure freedom 1 year after AEDs treatment. than those who received another treatment. There were Their study included 115 patients: 80% of them (92 large discrepancies in the frequency of use of certain cases) were enrolled in the training cohort and 20% (23 AEDs or their combinations between the model- cases) were enrolled in the validation cohort. Two hun- predicted AEDs regimens and the actually prescribed dred and seventy-nine candidate genes were involved regimens. The model performed even better than epilep- and five genes [rs658624 (SCN4B), rs678262 (SCN4B), tologists in clinical scenarios of monotherapy with leveti- rs2808526 (GABBR2), rs4869682 (SLC1A3), and racetam or lamotrigine. Regrettably, only 13% of the rs2283170 (KCNQ1)] were selected for the final model. actually prescribed AEDs regimens matched with the The model showed a good predictive accuracy of 83.5% regimen chosen by the model. Although this model was in the developmental cohort by cross-validation; its sen- based on a large sample size and was applied in clinical sitivity and positive predictive values were all above 80% practice, an obvious limitation of it is the lack of external in the two independent validation cohorts. However, the validation. Thus it still needs to be further optimized to sample size of this study was small, the external valid- improve the accuracy. An et al. [19] recently trained and ation was lacking, and the model was derived only from tested three algorithms, i.e. the multivariate logistic re- the traits of drug genes while not involving EEGs, MRIs, gression analysis, the support vector machine, and the or other key clinical and demographic characteristics, Yang et al. Acta Epileptologica (2021) 3:1 Page 4 of 6 which might affect its predictive performance. Shazadi the field of AEDs treatment, it is currently difficult to use et al. [21] assessed the validity of Petrovski’s algorithm this technology in clinical practice because of its complex- in two UK cohorts of newly diagnosed epilepsy patients, ity and the inconsistent variables. Specific software or web and showed that the multi-SNP prediction model was calculators need to be produced to facilitate clinical use not predictive for the initial treatment response. They and industrialization of the models. also found that the five SNPs appeared to have an im- pact on the prescription of carbamazepine or valproate Prediction models for the outcome of AEDs in the UK patients. withdrawal Some Chinese researchers have also investigated the use of About 70% of newly diagnosed PWEs could achieve seiz- machine learning techniques to predict AEDs effectiveness. ure freedom following appropriate AEDs therapy [12], Yao et al. [22] established five classical MLAs (decision tree, but the timing at which to stop AEDs is an important random forest, support vector machine, XGBoost, and logis- issue that remains a significant challenge for both pa- tic regression) to predict the outcomes of AEDs treatment in tients and doctors. Due to the fear of seizure relapse, patients with newly diagnosed epilepsy. They prospectively many PWEs choose to continue AEDs even after experi- collected information of 287 patients with newly diagnosed encing long-term seizure freedom, enduring side effects epilepsy and followed up the patients for a minimum of 3 of the treatment. If PWEs remain seizure free after AEDs years at the Second Affiliated Hospital of Zhejiang Univer- withdrawal, their psychological stress and quality of life sity. The patients were classified into the remission group can be significantly improved. In 2013, the Italian and non-remission group with regard to the outcome of League Against Epilepsy issued guidelines on AED with- seizure re-occurrence, and the former group was further di- drawal in PWEs who had achieved a long period of seiz- vided into the early remission group and late remission ure freedom [24], and these guidelines recommended group. The authors evaluated the performance of the models discontinuation of AEDs treatment after a minimum based on their precision, recall, F1-scores, and AUC values. seizure-free period of 2 years. It has been found that the The results showed that the XGBoost algorithm had the best earlier the drug is discontinued, the higher the chance of predictive performance between the remission group and seizure recurrence is. Some factors, such as abnormal non-remission group, with an F1 score of 0.947 and AUC EEGs, mental retardation, perinatal insults, abnormal value of 0.979, and between the early remission group and neurologic signs, partial seizures, older age of onset, and late remission group, with an F1 score of 0.836 and AUC female sex, can independently increase the risk of seizure value of 0.918. They claimed that the classified prediction relapse. Although the guidelines systematically evaluated could help doctors make clinical decisions and improve certain independent variables for AEDs withdrawal, there treatment strategies. In our previous study [23], we created a was not an integrated and comprehensive model for pre- model based on support vector machines (SVM) to predict dicting the outcome of AEDs withdrawal. the possibility of seizure freedom after levetiracetam therapy. In a retrospective study including 46 PWEs treated with leve- Statistical prediction models for AEDs withdrawal tiracetam, 80% of the patients were used to establish the In 2017, Lamberink et al. [25] established two nomo- SVM model and the other patients were used to subse- grams to predict the seizure recurrence and seizures in quently test the model. Before the start of levetiracetam the last year of follow-up after AEDs withdrawal in treatment, 11 clinical variables and four EEG parameters seizure-free patients. They initially did a systematic re- (sample entropies of α, β, θ, δ) were extracted. Our SVM view and meta-analysis to identify those studies, and model showed an accuracy of 72.2% in a five-fold cross- then they invited the author to participate in the re- validation, an accuracy of 75.0% in a jack-knife validation, search, including 1 769 PWEs with ten studies in the and an accuracy of 67.7% in a hold-out validation in the end. The adjusted concordance statistics were 0.65 for training cohort. The prediction accuracy of our model was predicting recurrence and 0.71 for predicting long-term 90% in the test cohort, and three different verification freedom; the calibration plots showed good performance methods all showed good reliability. The drawbacks of our of both models. This model also showed good perform- model were a lack of external validation and that the data ance of discrimination and a web-based calculator was were derived retrospectively from a single center; the sample subsequently built for practical purposes. The study had size was also small. Furthermore, the kernel function and di- a large sample size and the model was representative to mension of SVM could also have affected the accuracy of a certain extent and had some clinical value. Given that the model. Therefore, this model needs to be optimized and these nomograms were established from a pooled ana- performance must be improved by utilizing a larger dataset lysis of previously published data, the uniformity of clin- (Table 2). ical variables was slightly poor and multiple imputations While the development in machine learning technology were used to deal with the missing data. Furthermore, allows for more algorithms to be created and applied in there was a lack of external validation. Therefore, the Yang et al. Acta Epileptologica (2021) 3:1 Page 5 of 6 Table 2 Machine learning algorithms for the response to AEDs treatment Study Study Prediction target Training and Algorithms Predictive Performance External design testing cohorts features validation Devinsky Retrospective Choice of AEDs for 40, 000 patients for Machine learning About 5 000 AUC of 0.72 Yes et al. 2017 individual patients training and 10, 000 algorithm features [18] patients for testing An et al. Retrospective Prediction of DRE 175, 735 were training Multivariate logistic 1 270 features AUC of 0.76 No 2017 [19] cohort and the other regression, support 117, 157 were test vector machine and cohort random forest Petrovski Prospective Prediction of AEDs 115 patients with K-nearest neighbors 279 candidate Accuracy of 83.5% Yes et al. 2009 treatment outcomes newly diagnosed genes and sensitivity [20] epilepsy above 80% Yao et al. Retrospective Prediction of AEDs 287 patients with Decision tree, random Demographic F1 score and AUC No 2019 [22] treatment outcomes newly diagnosed forest, support vector features, medical value showed epilepsy machine, XGBoost and history, EEG and good performance logistic regression MRI Zhang et al. Retrospective Prediction of 46 patients with Support vector Clinical features 75.0% accuracy in No 2018 [23] efficacy of newly diagnosed machine and sample the training set levetiracetam epilepsy entropy and 90% in the test set AEDs Antiepileptic drugs, DRE Drug resistant epilepsy, AUC Area under the curve, EEG Electroencephalogram, MRI Magnetic resonance imaging universality of these nomograms requires further verifi- epilepsy surgery. They included 766 children from 15 cation by different research teams. Lin et al. [26] from centers, and the final models were composed of 3–5 the First Affiliated Hospital of Wenzhou Medical Uni- factors. The discrimination in terms of adjusted con- versity did an external validation of the Lamberink cordance statistic was 0.68 for predicting seizure re- model. The AUCs for predicting the recurrence and currence and 0.73 for predicting long-term seizure long-term outcomes were 0.71 and 0.68, respectively. freedom and the calibration plots showed good per- The calibration plots showed that the Lamberink two- formance. A visualized prediction tool is also pro- year model had a good fit and, with respect to the deci- vided online. In addition to the large sample size sion curve analysis, the Lamberink two-year model also from multiple centers, the validation of these nomo- had good performance. Lin’s research showed that the grams was executed well and supported by a web- Lamberink two-year model may have a greater value in based calculator. This indicates that these models guiding drug withdrawal in adult PWEs than other have high clinical value for recommending the cessa- models. tion and withholding of AEDs after pediatric epilepsy Lamberink et al. [27] also created two nomograms surgery. However, as there was no external validation, for individualized prediction of recurrence and long- the application of the nomograms in other popula- term outcomes of AEDs withdrawal after pediatric tions remains to be tested (Table 3). Table 3 Statistical prediction models for AEDs withdrawal Study Study Prediction target Development Final factors Validation Adjusted concordance-statistic Calibration design cohorts cohort Lamberink et al. Systematic Seizure relapse Ten studies Duration before remission, Internal and 0.65 for predicting seizure Calibration plots 2017 [25] review and and long-term with 1 769 seizure-free interval before external recurrence and 0.71 for showed good meta-analysis outcomes after patients AEDs withdrawal, age at validation predicting long-term calibration withdrawal of onset, history of febrile seizure freedom AEDs seizures, number of seizures before remission, epilepsy syndrome, developmental delay, EEG before withdrawal, sex, family history of epilepsy Lamberink et al. Retrospective Seizure relapse 766 children Age at withdrawal, time to Internal 0.68 for predicting seizure Calibration plots 2018 [27] and outcomes from 15 AEDs reduction, preoperative validation recurrence and 0.73 for showed good after AEDs European MRI, postoperative EEG, predicting eventual calibration withdrawal after epilepsy completeness of resection seizure freedom pediatric epilepsy centers of the anatomical lesion, surgery average frequency before surgery, number of AEDs at surgery AEDs Antiepileptic drugs, EEG Electroencephalogram, MRI Magnetic resonance imaging Yang et al. Acta Epileptologica (2021) 3:1 Page 6 of 6 Machine learning algorithm for the prognosis after AEDs 4. Naimo GD, Guarnaccia M, Sprovieri T, Ungaro C, Conforti FL, Andò S, et al. A systems biology approach for personalized medicine in withdrawal refractory epilepsy. Int J Mol Sci. 2019;20(15):1–15. There has been no report of MLA use for the prognosis 5. Shipe ME, Deppen SA, Farjah F, Grogan EL. Developing prediction or prediction of recurrence after AEDs withdrawal. Fu- models for clinical use using logistic regression: an overview. J Thorac Dis. 2019;11(Suppl 4):S574–S84. ture research is needed. 6. Han K, Song K, Choi BW. How to develop, validate, and compare clinical prediction models involving radiological parameters: study design and statistical methods. Korean J Radiol. 2016;17(3):339–50. Conclusion 7. D'Agostino RB, Vasan RS, Pencina MJ, Wolf PA, Cobain M, Massaro JM, et al. Several predictive models for AEDs regimen selection General cardiovascular risk profile for use in primary care: the Framingham Heart Study. Circulation. 2008;117(6):743–53. and withdrawal in epilepsy have been developed to fa- 8. Deo RC. Machine learning in medicine. Circulation. 2015;132(20):1920–30. cilitate clinical decisions, but there are some common 9. Chilamkurthy S, Ghosh R, Tanamala S, Biviji M, Campeau NG, Venugopal VK, issues caused by small sample size or inconsistent pa- et al. Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study. Lancet. 2018;392(10162):2388–96. rameters. In future studies, a group of main parame- 10. Coudray N, Ocampo PS, Sakellaropoulos T, Narula N, Snuderl M, Fenyö D, ters should be established initially. Then, more et al. Classification and mutation prediction from non-small cell lung cancer prospective, multi-center studies with large sample histopathology images using deep learning. Nat Med. 2018;24(10):1559–67. 11. Abbasi B, Goldenholz DM. Machine learning applications in epilepsy. sizes should be conducted to develop certain predict- Epilepsia. 2019;60(10):2037-47. ive models, which can be widely accepted in the field 12. Ben-Menachem E. Medical management of refractory epilepsy--practical of AEDs treatment in order to improve the prognosis treatment with novel antiepileptic drugs. Epilepsia. 2014;55(Suppl 1):3–8. 13. Perucca E, Tomson T. The pharmacological treatment of epilepsy in adults. and quality of life of patients with epilepsy. Lancet Neurol. 2011;10(5):446–56. 14. Walsh S, Donnan J, Fortin Y, Sikora L, Morrissey A, Collins K, et al. A Abbreviations systematic review of the risks factors associated with the onset and natural AEDs: Antiepileptic drugs; DRE: Drug refractory epilep\sy; PWE: Patients with progression of epilepsy. Neurotoxicology. 2017;61:64–77. epilepsy; MLA: Machine-learning algorithm; EEG: Electroencephalography; 15. Yang SJ,HeGN, HanX, WangN,Chen Y,Zhu XR,etal. Ascale for AUC: Area under the curve; SNP: Single nucleotide polymorphism; prediction of response to AEDs in patients with MRI-negative epilepsy. MRI: Magnetic resonance imaging; SVM: Support vector machines Epilepsy Behav. 2019;94:41–6. 16. Boonluksiri P, Visuthibhan A, Katanyuwong K. Clinical prediction rule of drug Acknowledgements resistant epilepsy in children. Epilepsy Res. 2015;5(2):84–8. None. 17. Tokatly Latzer I, Blumovich A, Sagi L, Uliel-Sibony S, Fattal-Valevski A. Prediction of drug-resistant epilepsy in children with cerebral palsy. J Child Authors’ contributions Neurol. 2020;35(3):187–94. SJY was responsible for study concept, literature review, manuscript 18. Devinsky O, Dilley C, Ozery-Flato M, Aharonov R, Goldschmidt Y, Rosen-Zvi composition and revision; BW conducted literature review. XH was M, et al. Changing the approach to treatment choice in epilepsy using big responsible for study concept and manuscript revision and obtained data. Epilepsy Behav. 2016;56:32–7. funding. All authors had read and approved the final manuscript. 19. An S, Malhotra K, Dilley C, Han-Burgess E, Valdez JN, Robertson J, et al. Predicting drug-resistant epilepsy—a machine learning approach based on administrative claims data. Epilepsy Behav. 2018;89:118–25. Funding 20. Petrovski S, Szoeke CE, Sheffield LJ, D'souza W, Huggins RM, O'brien TJ. This study was supported by Joint Construction Project of Province and Multi-SNP pharmacogenomic classifier is superior to single-SNP models Ministry in Henan Province (Grant number SB201901074). for predicting drug outcome in complex diseases. Pharmacogenet Genomics. 2009;19(2):147–52. Availability of data and materials 21. Shazadi K, Petrovski S, Roten A, Miller H, Huggins RM, Brodie MJ, et al. Not applicable. Validation of a multigenic model to predict seizure control in newly treated epilepsy. Epilepsy Res. 2014;108(10):1797–805. Ethics approval and consent to participate 22. Yao L, Cai M, Chen Y, Shen C, Shi L, Guo Y, et al. Prediction of antiepileptic Not applicable. drug treatment outcomes of patients with newly diagnosed epilepsy by machine learning. Epilepsy Behav. 2019;96:92–7. 23. Zhang JH, Han X, Zhao HW, Zhao D, Wang N, Zhao T, et al. Personalized Consent for publication prediction model for seizure-free epilepsy with levetiracetam therapy: a The author gives consent for publication. retrospective data analysis using support vector machine. Br J Clin Pharmacol. 2018;84(11):2615–24. Competing interests 24. BeghiE,GiussaniG,GrossoS,IudiceA,LaNeveA,PisaniF,et al. The authors declare that they have no conflict of interest. Withdrawal of antiepileptic drugs: guidelines of the Italian League Against Epilepsy. Epilepsia. 2013;54(Suppl 7):2–12. Received: 27 May 2020 Accepted: 13 November 2020 25. Lamberink HJ, Otte WM, Geerts AT, Pavlovic M, Ramos-Lizana J, Verweg J, et al. Individualised prediction model of seizure recurrence and long- term outcomes after withdrawal of antiepileptic drugs in seizure-free References patients: a systematic review and individual participant data meta- 1. Ngugi AK, Bottomley C, Kleinschmidt I, Sander JW, Newton CR. analysis. Lancet Neurol. 2017;16(7):523–31. Estimation of the burden of active and life-time epilepsy: a meta- 26. Pisani F, Facini C, Pavlidis E, Spagnoli C, Boylan G. Epilepsy after neonatal analytic approach. Epilepsia. 2010;51(5):883–90. seizures: literature review. Eur J Paediatr Neurol. 2015;19(1):6–14. 2. Chen Z, Brodie MJ, Liew D, Kwan P. Treatment outcomes in patients with 27. Lamberink HJ, Boshuisen K, Otte WM, Geleijns K, Braun KPJ. Individualized newly diagnosed epilepsy treated with established and new antiepileptic prediction of seizure relapse and outcomes following antiepileptic drug drugs: a 30-year longitudinal cohort study. JAMA Neurol. 2018;75(3):279–86. withdrawal after pediatric epilepsy surgery. Epilepsia. 2018;59(3):e28–33. 3. Kwan P, Arzimanoglou A, Berg AT, Brodie MJ, Allen Hauser W, Mathern G, et al. Definition of drug resistant epilepsy: consensus proposal by the ad hoc Task Force of the ILAE Commission on Therapeutic Strategies. Epilepsia. 2010;51(6):1069–77. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Acta Epileptologica Springer Journals

Models for predicting treatment efficacy of antiepileptic drugs and prognosis of treatment withdrawal in epilepsy patients

Acta Epileptologica , Volume 3 (1) – Jan 11, 2021

Loading next page...
 
/lp/springer-journals/models-for-predicting-treatment-efficacy-of-antiepileptic-drugs-and-6YLzzRMN3n

References (28)

Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2021
eISSN
2524-4434
DOI
10.1186/s42494-020-00035-9
Publisher site
See Article on Publisher Site

Abstract

Although antiepileptic drugs (AEDs) are the most effective treatment for epilepsy, 30–40% of patients with epilepsy would develop drug-refractory epilepsy. An accurate, preliminary prediction of the efficacy of AEDs has great clinical significance for patient treatment and prognosis. Some studies have developed statistical models and machine- learning algorithms (MLAs) to predict the efficacy of AEDs treatment and the progression of disease after treatment withdrawal, in order to provide assistance for making clinical decisions in the aim of precise, personalized treatment. The field of prediction models with statistical models and MLAs is attracting growing interest and is developing rapidly. What’s more, more and more studies focus on the external validation of the existing model. In this review, we will give a brief overview of recent developments in this discipline. Keywords: Prediction model, Machine learning, Antiepileptic drugs, Drug response, Withdrawal reaction Background burden on their families. Accurate prediction of the effi- Epilepsy is one of the most common neurological dis- cacy of AEDs before the initiation of treatment can reduce eases and has affected more than 68 million people the use of ineffective drugs, alleviate patients’ pain, and worldwide [1]. Although antiepileptic drugs (AEDs) are improve the prognosis in patients. currently the primary treatment option for patients with Doctors and patients often find it hard to decide epilepsy (PWE), about 40% of PWEs will suffer the con- whether to reduce or stop AEDs use. Although PWEs sequence of drug-resistant epilepsy (DRE) [2]. According can have better cognitive performances and higher qual- to the International League Against Epilepsy, DRE is de- ity of life after AEDs withdrawal, there is also an in- fined as the failure to achieve sustained seizure freedom creased risk of recurrence. To avoid recurrence, many after adequate trials of two tolerated and appropriately patients decide to put up with the side effects rather chosen AEDs treatments (monotherapies or combination than completely withdrawing the AEDs. Therefore, there therapies) [3]. The mechanism of DRE is not fully clear is an urgent need of effective and practical decision- and may be related to the sensitivity of drug targets, activ- making tools to assist clinicians to establish the course ity of drug transporters, cytochrome P450, structural of AEDs treatment and withdrawal, as well as to help neural network, and other potential causes of epilepsy [4]. realize precise, personalized treatment. Abrupt and repetitive seizures may lead to neurobiochem- Prediction models can integrate multiple clinical or ical changes in the brain, cognitive decline, and serious non-clinical parameters within a certain time to calcu- psychological problems in PWEs, which can seriously late the probability of diagnostic outcomes as well as the affect patients’ quality of life and cause an increased disease prognosis. These models can stratify patient risk stratification to support clinical decision-making and im- prove the prognosis and quality of care for patients [5]. * Correspondence: hanxiong7589@126.com Prediction models are divided into two main categories: Department of Neurology, People’s Hospital of Zhengzhou University, Henan Provincial People’s Hospital, Zhengzhou 450003, China © The Author(s). 2021 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Yang et al. Acta Epileptologica (2021) 3:1 Page 2 of 6 those based on statistics and those based on machine comorbidities, history of central nervous system infec- learning algorithms (MLAs). tion, cognitive impairment, epilepsy syndrome, presence The statistical prediction models are those whose de- of structural abnormalities in magnetic resonance im- velopment is based on statistics, such as univariate and aging (MRI), previous history of status epilepticus, family multivariate logistic regression/COX regression analysis history of epilepsy, history of perinatal brain injury, and to select prediction variables. The integration of multiple certain electroencephalography (EEG) features [12–15]. selection variables is used to calculate the probability of The identification and determination of these parame- a particular diagnosis or disease prognosis [5]. The de- ters are the basis for creating a predictive model. velopment of statistical prediction models involves col- lection of datasets, selection of prediction variables, Statistical prediction models development of a prediction model, evaluation of the Based on the statistical method, certain variables can be model’s performance, internal and external validation, selected and integrated into a model and a score system and further update of the model. This type of method can be created for a specific purpose; this type of model can be used to create an easy-to-use prediction scoring has been tested in the field of epilepsy. Boonluksiri et al. system [6]. One example is the Framingham risk score, [16] enrolled 308 children with epilepsy in a retrospect- which is widely used in the public health field for esti- ive study, and they selected the age at onset, prior mating the probability of the occurrence of cardiovascu- neurological deficits, and abnormal EEGs as variables, lar diseases in an individual within the next 10 years. and established a scale for predicting DRE in children. This model was built based on the traditional prediction The children were then divided into 3 groups depending variables of age, sex, systolic blood pressure, hyperten- on the risk of developing DRE: low risk (score < 6 sion treatment, total and high-density lipoprotein chol- points), moderate risk (score 6–12 points) and high risk esterol levels, smoking, and diabetes [7]. (score > 12 points), with positive likelihood ratios of 0.5, With the development of artificial intelligence, ma- 1.8, and 12.5, respectively, and an area under the curve chine learning is concerned with algorithm induction to (AUC) of 0.76. However, as this retrospective study was improve model performance by using statistical and conducted in a single center with a small sample size computer science approaches, and has shown potentials and lacked both internal and external validation, the per- for industrialization. Machine learning has also been ap- formance and practicality of this model need further val- plied to fields like speech recognition, image classifica- idation. In a previous study [15], we developed a scale tion, text translation, and medical care [8] for the for predicting DRE in adult patients with MRI-negative detection of critical findings in head computerized tom- epilepsy [MRI(−)DRE]. The AUC was 0.89 and the risk ography scans [9] and the classification of cancer [10]. stratification was given as: low risk (0–3 points), medium The machine learning technique is superior to manual as- risk (3–5 points), and high risk (> 5 points). Using this sessment by clinical experts in that it has higher accuracy method, the probability of DRE could also be calculated. of diagnosis and outcome prediction, and it can also be However, this scale was limited by the retrospective de- used for epilepsy, especially for automated seizure detec- sign based on data of 132 patients, so further validation tion, analysis of imaging and clinical data, epilepsy is needed. Latzer et al. [17] created a model to predict localization, and prediction of medical and surgical out- DRE in children with cerebral palsy at the Tel Aviv comes [11]. Additional validation techniques, such as the Medical Center in Tel Aviv, Israel and this model was hold-out cross-validation, k-fold cross validation, and used in a retrospective study including 118 patients. The “leave-one-out method” cross-validation, can be used to es- model was composed of four parameters (low Apgar timate the performance of the technique. In the following, score at 5 min, neonatal seizures, focal-onset epilepsy, we will give a brief overview of the efficacy of statistical pre- and focal slowing on EEG) and the AUC was 0.84. Al- diction models and MLAs for predicting AEDs treatment though their model helped to identify which patient response and patients’ outcome after AEDs withdrawal. would achieve better seizure control, the study was per- formed with a small sample size and the lack of valid- Prediction models for the response to AEDs ation makes it difficult to judge the model’s performance treatment (Table 1). Drug selection mainly relies on official guidelines and In summary, a few comprehensive statistical models clinical experience of doctors, due to the fact that the with multiple variables have been established for predict- treatment efficacy varies among individual PWEs. Per- ing the response to AEDs treatment, but they were sonalized selection of effective AEDs still remains a big weakened by some limitations including the retrospect- challenge. Some prospective studies have identified cer- ive design, small sample sizes, and the lack of internal tain predictive factors of DRE, such as early onset, sex, and external validation. To address these, more pro- duration of epilepsy, multiple seizure types, spective, multi-center studies with large sample sizes are Yang et al. Acta Epileptologica (2021) 3:1 Page 3 of 6 Table 1 Statistical prediction models for the response to AEDs treatment Study Study Prediction Development Final factors Validation AUC Stratified risks design target cohort cohort Boonluksiri Retrospective DRE in children 308 cases in Hatyai Age onset, prior neurological No 0.76 low risk: < 6 points; et al. 2015 [16] Hospital in Thailand deficits and abnormal EEG moderate risk: 6–12 points; high risk: > 12 points Yang et al. Retrospective DRE in adults 132 cases at Henan EEG before AEDs, history of No 0.89 low risk: < 3 points; 2019 [15] Provincial People’s CNS infection, initial precipitating moderate risk: 3–5 Hospital in China injuries, and more than one points; high risk: > 5 recurrence in the first 6 months points Latzer et al. Retrospective DRE in children 281 children with Low Apgar score at 5 min, No 0.68 No 2019 [17] with cerebral cerebral palsy at the neonatal seizures, focal-onset palsy Dana-Dwek Children’s epilepsy and focal slowing on EEG Hospital DRE Drug resistant epilepsy, AUC Area under the curve, EEG Electroencephalogram, MRI Magnetic resonance imaging, CNS Central nervous system required. In addition, these models need to be verified in random forest algorithm, to identify patients at high risk multiple centers. of DRE. A total of 292, 892 patients met the inclusion criteria for epilepsy; 175, 735 of them were assigned to Machine learning algorithms the training cohort and the other 117, 157 were assigned MLAs can be used to extract more EEG, imaging, and to the test cohort, and 1 270 features were screened as clinical features of patients to build prediction models predictive factors. The random forest algorithm had an and validate performance through the use of more AUC of 0.76 and performed the best of the three methods. MLAs can also be readily applied in artificial models. It could predict the emergence of DRE approxi- intelligence-based industrialization, an extremely rele- mately 2 years in advance before a patient failed two vant and competitive field today. AEDs trials. The drawback of this study was that it was UCB Pharma has been actively involved in conducting a retrospective study without external validation. Fur- research on the development and validation of MLA for thermore, the DRE incidence was only 13.1%, which was use in the prediction of AEDs effectiveness in individual lower than that in other studies, indicating that the data- PWEs. Devinsky et al. [18] at the New York University set had significant limitations. Medical Center, based on the UCB–IBM collaboration, A number of pharmacogenomic studies have focused explored the application of MLA to construct an algo- on identifying single nucleotide polymorphism (SNP) rithm for AEDs prescription. A total of 50, 000 PWEs markers for predicting the outcomes of AEDs treat- were retrospectively enrolled in the study and randomly ments, and some studies have tried to establish certain divided into a training group of 40, 000 patients and a multi-SNP models to predict the response to AEDs. Pet- testing group of 10, 000 patients. Roughly 5, 000 features rovski et al. [20] prospectively collected the genetic re- were extracted to build the prediction model, which had sults of patients with newly diagnosed epilepsy and an AUC of 0.72 and was considered to have a good pre- developed a multi-SNP classification model, based on dictive power. The patients with the model-predicted the k-nearest neighbor supervised learning approach, to AEDs regimen had significantly higher survival rates predict the seizure freedom 1 year after AEDs treatment. than those who received another treatment. There were Their study included 115 patients: 80% of them (92 large discrepancies in the frequency of use of certain cases) were enrolled in the training cohort and 20% (23 AEDs or their combinations between the model- cases) were enrolled in the validation cohort. Two hun- predicted AEDs regimens and the actually prescribed dred and seventy-nine candidate genes were involved regimens. The model performed even better than epilep- and five genes [rs658624 (SCN4B), rs678262 (SCN4B), tologists in clinical scenarios of monotherapy with leveti- rs2808526 (GABBR2), rs4869682 (SLC1A3), and racetam or lamotrigine. Regrettably, only 13% of the rs2283170 (KCNQ1)] were selected for the final model. actually prescribed AEDs regimens matched with the The model showed a good predictive accuracy of 83.5% regimen chosen by the model. Although this model was in the developmental cohort by cross-validation; its sen- based on a large sample size and was applied in clinical sitivity and positive predictive values were all above 80% practice, an obvious limitation of it is the lack of external in the two independent validation cohorts. However, the validation. Thus it still needs to be further optimized to sample size of this study was small, the external valid- improve the accuracy. An et al. [19] recently trained and ation was lacking, and the model was derived only from tested three algorithms, i.e. the multivariate logistic re- the traits of drug genes while not involving EEGs, MRIs, gression analysis, the support vector machine, and the or other key clinical and demographic characteristics, Yang et al. Acta Epileptologica (2021) 3:1 Page 4 of 6 which might affect its predictive performance. Shazadi the field of AEDs treatment, it is currently difficult to use et al. [21] assessed the validity of Petrovski’s algorithm this technology in clinical practice because of its complex- in two UK cohorts of newly diagnosed epilepsy patients, ity and the inconsistent variables. Specific software or web and showed that the multi-SNP prediction model was calculators need to be produced to facilitate clinical use not predictive for the initial treatment response. They and industrialization of the models. also found that the five SNPs appeared to have an im- pact on the prescription of carbamazepine or valproate Prediction models for the outcome of AEDs in the UK patients. withdrawal Some Chinese researchers have also investigated the use of About 70% of newly diagnosed PWEs could achieve seiz- machine learning techniques to predict AEDs effectiveness. ure freedom following appropriate AEDs therapy [12], Yao et al. [22] established five classical MLAs (decision tree, but the timing at which to stop AEDs is an important random forest, support vector machine, XGBoost, and logis- issue that remains a significant challenge for both pa- tic regression) to predict the outcomes of AEDs treatment in tients and doctors. Due to the fear of seizure relapse, patients with newly diagnosed epilepsy. They prospectively many PWEs choose to continue AEDs even after experi- collected information of 287 patients with newly diagnosed encing long-term seizure freedom, enduring side effects epilepsy and followed up the patients for a minimum of 3 of the treatment. If PWEs remain seizure free after AEDs years at the Second Affiliated Hospital of Zhejiang Univer- withdrawal, their psychological stress and quality of life sity. The patients were classified into the remission group can be significantly improved. In 2013, the Italian and non-remission group with regard to the outcome of League Against Epilepsy issued guidelines on AED with- seizure re-occurrence, and the former group was further di- drawal in PWEs who had achieved a long period of seiz- vided into the early remission group and late remission ure freedom [24], and these guidelines recommended group. The authors evaluated the performance of the models discontinuation of AEDs treatment after a minimum based on their precision, recall, F1-scores, and AUC values. seizure-free period of 2 years. It has been found that the The results showed that the XGBoost algorithm had the best earlier the drug is discontinued, the higher the chance of predictive performance between the remission group and seizure recurrence is. Some factors, such as abnormal non-remission group, with an F1 score of 0.947 and AUC EEGs, mental retardation, perinatal insults, abnormal value of 0.979, and between the early remission group and neurologic signs, partial seizures, older age of onset, and late remission group, with an F1 score of 0.836 and AUC female sex, can independently increase the risk of seizure value of 0.918. They claimed that the classified prediction relapse. Although the guidelines systematically evaluated could help doctors make clinical decisions and improve certain independent variables for AEDs withdrawal, there treatment strategies. In our previous study [23], we created a was not an integrated and comprehensive model for pre- model based on support vector machines (SVM) to predict dicting the outcome of AEDs withdrawal. the possibility of seizure freedom after levetiracetam therapy. In a retrospective study including 46 PWEs treated with leve- Statistical prediction models for AEDs withdrawal tiracetam, 80% of the patients were used to establish the In 2017, Lamberink et al. [25] established two nomo- SVM model and the other patients were used to subse- grams to predict the seizure recurrence and seizures in quently test the model. Before the start of levetiracetam the last year of follow-up after AEDs withdrawal in treatment, 11 clinical variables and four EEG parameters seizure-free patients. They initially did a systematic re- (sample entropies of α, β, θ, δ) were extracted. Our SVM view and meta-analysis to identify those studies, and model showed an accuracy of 72.2% in a five-fold cross- then they invited the author to participate in the re- validation, an accuracy of 75.0% in a jack-knife validation, search, including 1 769 PWEs with ten studies in the and an accuracy of 67.7% in a hold-out validation in the end. The adjusted concordance statistics were 0.65 for training cohort. The prediction accuracy of our model was predicting recurrence and 0.71 for predicting long-term 90% in the test cohort, and three different verification freedom; the calibration plots showed good performance methods all showed good reliability. The drawbacks of our of both models. This model also showed good perform- model were a lack of external validation and that the data ance of discrimination and a web-based calculator was were derived retrospectively from a single center; the sample subsequently built for practical purposes. The study had size was also small. Furthermore, the kernel function and di- a large sample size and the model was representative to mension of SVM could also have affected the accuracy of a certain extent and had some clinical value. Given that the model. Therefore, this model needs to be optimized and these nomograms were established from a pooled ana- performance must be improved by utilizing a larger dataset lysis of previously published data, the uniformity of clin- (Table 2). ical variables was slightly poor and multiple imputations While the development in machine learning technology were used to deal with the missing data. Furthermore, allows for more algorithms to be created and applied in there was a lack of external validation. Therefore, the Yang et al. Acta Epileptologica (2021) 3:1 Page 5 of 6 Table 2 Machine learning algorithms for the response to AEDs treatment Study Study Prediction target Training and Algorithms Predictive Performance External design testing cohorts features validation Devinsky Retrospective Choice of AEDs for 40, 000 patients for Machine learning About 5 000 AUC of 0.72 Yes et al. 2017 individual patients training and 10, 000 algorithm features [18] patients for testing An et al. Retrospective Prediction of DRE 175, 735 were training Multivariate logistic 1 270 features AUC of 0.76 No 2017 [19] cohort and the other regression, support 117, 157 were test vector machine and cohort random forest Petrovski Prospective Prediction of AEDs 115 patients with K-nearest neighbors 279 candidate Accuracy of 83.5% Yes et al. 2009 treatment outcomes newly diagnosed genes and sensitivity [20] epilepsy above 80% Yao et al. Retrospective Prediction of AEDs 287 patients with Decision tree, random Demographic F1 score and AUC No 2019 [22] treatment outcomes newly diagnosed forest, support vector features, medical value showed epilepsy machine, XGBoost and history, EEG and good performance logistic regression MRI Zhang et al. Retrospective Prediction of 46 patients with Support vector Clinical features 75.0% accuracy in No 2018 [23] efficacy of newly diagnosed machine and sample the training set levetiracetam epilepsy entropy and 90% in the test set AEDs Antiepileptic drugs, DRE Drug resistant epilepsy, AUC Area under the curve, EEG Electroencephalogram, MRI Magnetic resonance imaging universality of these nomograms requires further verifi- epilepsy surgery. They included 766 children from 15 cation by different research teams. Lin et al. [26] from centers, and the final models were composed of 3–5 the First Affiliated Hospital of Wenzhou Medical Uni- factors. The discrimination in terms of adjusted con- versity did an external validation of the Lamberink cordance statistic was 0.68 for predicting seizure re- model. The AUCs for predicting the recurrence and currence and 0.73 for predicting long-term seizure long-term outcomes were 0.71 and 0.68, respectively. freedom and the calibration plots showed good per- The calibration plots showed that the Lamberink two- formance. A visualized prediction tool is also pro- year model had a good fit and, with respect to the deci- vided online. In addition to the large sample size sion curve analysis, the Lamberink two-year model also from multiple centers, the validation of these nomo- had good performance. Lin’s research showed that the grams was executed well and supported by a web- Lamberink two-year model may have a greater value in based calculator. This indicates that these models guiding drug withdrawal in adult PWEs than other have high clinical value for recommending the cessa- models. tion and withholding of AEDs after pediatric epilepsy Lamberink et al. [27] also created two nomograms surgery. However, as there was no external validation, for individualized prediction of recurrence and long- the application of the nomograms in other popula- term outcomes of AEDs withdrawal after pediatric tions remains to be tested (Table 3). Table 3 Statistical prediction models for AEDs withdrawal Study Study Prediction target Development Final factors Validation Adjusted concordance-statistic Calibration design cohorts cohort Lamberink et al. Systematic Seizure relapse Ten studies Duration before remission, Internal and 0.65 for predicting seizure Calibration plots 2017 [25] review and and long-term with 1 769 seizure-free interval before external recurrence and 0.71 for showed good meta-analysis outcomes after patients AEDs withdrawal, age at validation predicting long-term calibration withdrawal of onset, history of febrile seizure freedom AEDs seizures, number of seizures before remission, epilepsy syndrome, developmental delay, EEG before withdrawal, sex, family history of epilepsy Lamberink et al. Retrospective Seizure relapse 766 children Age at withdrawal, time to Internal 0.68 for predicting seizure Calibration plots 2018 [27] and outcomes from 15 AEDs reduction, preoperative validation recurrence and 0.73 for showed good after AEDs European MRI, postoperative EEG, predicting eventual calibration withdrawal after epilepsy completeness of resection seizure freedom pediatric epilepsy centers of the anatomical lesion, surgery average frequency before surgery, number of AEDs at surgery AEDs Antiepileptic drugs, EEG Electroencephalogram, MRI Magnetic resonance imaging Yang et al. Acta Epileptologica (2021) 3:1 Page 6 of 6 Machine learning algorithm for the prognosis after AEDs 4. Naimo GD, Guarnaccia M, Sprovieri T, Ungaro C, Conforti FL, Andò S, et al. A systems biology approach for personalized medicine in withdrawal refractory epilepsy. Int J Mol Sci. 2019;20(15):1–15. There has been no report of MLA use for the prognosis 5. Shipe ME, Deppen SA, Farjah F, Grogan EL. Developing prediction or prediction of recurrence after AEDs withdrawal. Fu- models for clinical use using logistic regression: an overview. J Thorac Dis. 2019;11(Suppl 4):S574–S84. ture research is needed. 6. Han K, Song K, Choi BW. How to develop, validate, and compare clinical prediction models involving radiological parameters: study design and statistical methods. Korean J Radiol. 2016;17(3):339–50. Conclusion 7. D'Agostino RB, Vasan RS, Pencina MJ, Wolf PA, Cobain M, Massaro JM, et al. Several predictive models for AEDs regimen selection General cardiovascular risk profile for use in primary care: the Framingham Heart Study. Circulation. 2008;117(6):743–53. and withdrawal in epilepsy have been developed to fa- 8. Deo RC. Machine learning in medicine. Circulation. 2015;132(20):1920–30. cilitate clinical decisions, but there are some common 9. Chilamkurthy S, Ghosh R, Tanamala S, Biviji M, Campeau NG, Venugopal VK, issues caused by small sample size or inconsistent pa- et al. Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study. Lancet. 2018;392(10162):2388–96. rameters. In future studies, a group of main parame- 10. Coudray N, Ocampo PS, Sakellaropoulos T, Narula N, Snuderl M, Fenyö D, ters should be established initially. Then, more et al. Classification and mutation prediction from non-small cell lung cancer prospective, multi-center studies with large sample histopathology images using deep learning. Nat Med. 2018;24(10):1559–67. 11. Abbasi B, Goldenholz DM. Machine learning applications in epilepsy. sizes should be conducted to develop certain predict- Epilepsia. 2019;60(10):2037-47. ive models, which can be widely accepted in the field 12. Ben-Menachem E. Medical management of refractory epilepsy--practical of AEDs treatment in order to improve the prognosis treatment with novel antiepileptic drugs. Epilepsia. 2014;55(Suppl 1):3–8. 13. Perucca E, Tomson T. The pharmacological treatment of epilepsy in adults. and quality of life of patients with epilepsy. Lancet Neurol. 2011;10(5):446–56. 14. Walsh S, Donnan J, Fortin Y, Sikora L, Morrissey A, Collins K, et al. A Abbreviations systematic review of the risks factors associated with the onset and natural AEDs: Antiepileptic drugs; DRE: Drug refractory epilep\sy; PWE: Patients with progression of epilepsy. Neurotoxicology. 2017;61:64–77. epilepsy; MLA: Machine-learning algorithm; EEG: Electroencephalography; 15. Yang SJ,HeGN, HanX, WangN,Chen Y,Zhu XR,etal. Ascale for AUC: Area under the curve; SNP: Single nucleotide polymorphism; prediction of response to AEDs in patients with MRI-negative epilepsy. MRI: Magnetic resonance imaging; SVM: Support vector machines Epilepsy Behav. 2019;94:41–6. 16. Boonluksiri P, Visuthibhan A, Katanyuwong K. Clinical prediction rule of drug Acknowledgements resistant epilepsy in children. Epilepsy Res. 2015;5(2):84–8. None. 17. Tokatly Latzer I, Blumovich A, Sagi L, Uliel-Sibony S, Fattal-Valevski A. Prediction of drug-resistant epilepsy in children with cerebral palsy. J Child Authors’ contributions Neurol. 2020;35(3):187–94. SJY was responsible for study concept, literature review, manuscript 18. Devinsky O, Dilley C, Ozery-Flato M, Aharonov R, Goldschmidt Y, Rosen-Zvi composition and revision; BW conducted literature review. XH was M, et al. Changing the approach to treatment choice in epilepsy using big responsible for study concept and manuscript revision and obtained data. Epilepsy Behav. 2016;56:32–7. funding. All authors had read and approved the final manuscript. 19. An S, Malhotra K, Dilley C, Han-Burgess E, Valdez JN, Robertson J, et al. Predicting drug-resistant epilepsy—a machine learning approach based on administrative claims data. Epilepsy Behav. 2018;89:118–25. Funding 20. Petrovski S, Szoeke CE, Sheffield LJ, D'souza W, Huggins RM, O'brien TJ. This study was supported by Joint Construction Project of Province and Multi-SNP pharmacogenomic classifier is superior to single-SNP models Ministry in Henan Province (Grant number SB201901074). for predicting drug outcome in complex diseases. Pharmacogenet Genomics. 2009;19(2):147–52. Availability of data and materials 21. Shazadi K, Petrovski S, Roten A, Miller H, Huggins RM, Brodie MJ, et al. Not applicable. Validation of a multigenic model to predict seizure control in newly treated epilepsy. Epilepsy Res. 2014;108(10):1797–805. Ethics approval and consent to participate 22. Yao L, Cai M, Chen Y, Shen C, Shi L, Guo Y, et al. Prediction of antiepileptic Not applicable. drug treatment outcomes of patients with newly diagnosed epilepsy by machine learning. Epilepsy Behav. 2019;96:92–7. 23. Zhang JH, Han X, Zhao HW, Zhao D, Wang N, Zhao T, et al. Personalized Consent for publication prediction model for seizure-free epilepsy with levetiracetam therapy: a The author gives consent for publication. retrospective data analysis using support vector machine. Br J Clin Pharmacol. 2018;84(11):2615–24. Competing interests 24. BeghiE,GiussaniG,GrossoS,IudiceA,LaNeveA,PisaniF,et al. The authors declare that they have no conflict of interest. Withdrawal of antiepileptic drugs: guidelines of the Italian League Against Epilepsy. Epilepsia. 2013;54(Suppl 7):2–12. Received: 27 May 2020 Accepted: 13 November 2020 25. Lamberink HJ, Otte WM, Geerts AT, Pavlovic M, Ramos-Lizana J, Verweg J, et al. Individualised prediction model of seizure recurrence and long- term outcomes after withdrawal of antiepileptic drugs in seizure-free References patients: a systematic review and individual participant data meta- 1. Ngugi AK, Bottomley C, Kleinschmidt I, Sander JW, Newton CR. analysis. Lancet Neurol. 2017;16(7):523–31. Estimation of the burden of active and life-time epilepsy: a meta- 26. Pisani F, Facini C, Pavlidis E, Spagnoli C, Boylan G. Epilepsy after neonatal analytic approach. Epilepsia. 2010;51(5):883–90. seizures: literature review. Eur J Paediatr Neurol. 2015;19(1):6–14. 2. Chen Z, Brodie MJ, Liew D, Kwan P. Treatment outcomes in patients with 27. Lamberink HJ, Boshuisen K, Otte WM, Geleijns K, Braun KPJ. Individualized newly diagnosed epilepsy treated with established and new antiepileptic prediction of seizure relapse and outcomes following antiepileptic drug drugs: a 30-year longitudinal cohort study. JAMA Neurol. 2018;75(3):279–86. withdrawal after pediatric epilepsy surgery. Epilepsia. 2018;59(3):e28–33. 3. Kwan P, Arzimanoglou A, Berg AT, Brodie MJ, Allen Hauser W, Mathern G, et al. Definition of drug resistant epilepsy: consensus proposal by the ad hoc Task Force of the ILAE Commission on Therapeutic Strategies. Epilepsia. 2010;51(6):1069–77.

Journal

Acta EpileptologicaSpringer Journals

Published: Jan 11, 2021

There are no references for this article.