Authors :
Presenting Author: Chantelle Lin, HBSc – University of Calgary
Simeon Platte, MS – Goethe-Universität Frankfurt am Main; Jordan Engbers, PhD – University of Calgary; Chantal Depondt, MD PhD – Université Libre de Bruxelles; Sophie von Brauchitsch, MD – Goethe-Universität Frankfurt am Main; Felix Rosenow, MD – Goethe-Universität Frankfurt am Main; Reetta Kälviäinen, MD PhD – University of Eastern Finland; Giorgia Guerini, MD PhD – Université Libre de Bruxelles; Afsheen Kumar, PhD – Goethe-Universität Frankfurt am Main; Samuel Wiebe, MD MSc – University of Calgary; Amy Brooks-Kayal, MD – University of California, Davis; Spiros Denaxas, PhD – University College London; Roland Krause, PhD – Université du Luxembourg; Tolulope Sajobi, PhD – University of Calgary; Nils Forkert, PhD – University of Calgary; Collaborators Epi25, Consortium – Epi25; Massimo Pandolfo, MD – McGill University; Andreas Chiocchetti, PhD – Goethe-Universität Frankfurt am Main; Karl Klein, MD PhD – University of Calgary; Colin Josephson, MD MSc – University of Calgary
Rationale:
With over 20 anti-seizure medications (ASMs) available, identifying the ideal drug is often imprecise and time-consuming. Developing predictive models to expedite optimal drug selection is challenging due to the minimal differences in efficacy. However, side-effects vary considerably between medications, and are one of the main reasons for discontinuation of ASM treatment. We combined clinical and genetic data to train classification algorithms to predict if patients will develop side-effects from valproic acid (VPA) treatment to assist physicians in prescribing the optimal ASM for their patients.
Methods:
This retrospective cohort study included 624 patients exposed to VPA from the RAISE-GENIC study during the years 2009 to 2020. The predicted outcome was defined as ASM discontinuation due to any side-effect. Clinical features included age of onset, patient age, sex, comorbidities, seizure type, EEG variables, and imaging variables. Network analysis of mRNA expression data from VPA-exposed neurons derived from control induced pluripotent stem cells (iPSCs) was leveraged to extract features from exome sequencing and genome-wide single nucleotide polymorphism data. Features were selected for model inclusion through a ReliefF algorithm. Penalized logistic regression, support vector machine, random forest, and k-nearest neighbor models were trained on a balanced normalized bootstrapped training dataset and model following repeated stratified 10-fold cross validation. Model discrimination and calibration were used for quantitative evaluation. Models with only clinical and combined clinical and genetic features were compared.
Results:
Of the 624 patients, 352 (56.4%) were females, and median age of onset after imputation and patient’s age at data collection were 16 (interquartile range [IQR] 9-27) and 46 (IQR 34-59) years, respectively. There were 112 (17.9%) patients that stopped VPA due to side-effects. A total of 15 clinical features and 485 genetic features were included following ReliefF selection. The best performing model was penalized logistic regression that included both genetic and clinical features. Area under the receiver operating characteristic curve (AUC) was 0.66 [0.66-0.67], Brier score was 0.20 [0.19-0.21], sensitivity was 0.42 [0.41-0.0.42], and specificity 0.82 [0.82-0.83].
Conclusions:
Machine learning using clinical and genetic features can predict treatment-ending side-effects to VPA with moderate performance, discrimination, and calibration. If these models can be refined and externally validated, decision tools can be incorporated into clinical routines, simplifying drug prescriptions, saving time, and improving patient quality of life.
Funding:
The RAISE-GENIC consortium is funded by ERA PerMed, an ERA-Net Cofund, the Canadian Institutes of Health Research, the Federal Ministry of Education and Research, Germany, Academy of Finland, and Fund for Scientific Research, Belgium.