Wielomodałowy model sztucznej inteligencji przewidujący odpowiedź na leczenie epilepsji lekiem walproinian
A translational multimodal machine-learning prototype predicting valproate response in epilepsy treatment
W skrócie
Naukowcy stworzyli narzędzie oparte na sztucznej inteligencji, które analizuje dane genetyczne, komórkowe i kliniczne pacjentów, aby przewidzieć, czy lek walproinian będzie skuteczny w leczeniu epilepsji. W badaniu przeprowadzonym na ponad 400 pacjentach model osiągnął dokładność około 63 procent i poprawnie zidentyfikował 70 procent pacjentów, u których lek by nie zadziałał. Chociaż narzędzie nie jest jeszcze gotowe do codziennego stosowania w szpitalach, wykazuje obiecujący potencjał do pomocy lekarzom w wyborze najlepszego leku dla każdego pacjenta i szybszym znalezieniu leczenia, które będzie skuteczne.
Oryginalny abstract (angielski)
OBJECTIVE: Epilepsy affects ~1% of the global population and often requires lifelong antiseizure medication (ASM) therapy. Valproic acid (VPA) is a commonly prescribed first-line ASM, yet only approximately half of patients achieve sustained seizure freedom. Treatment selection remains largely empirical. We aimed to develop and independently validate a multimodal predictive model to estimate response to VPA and support more individualized treatment strategies. METHODS: This cross-sectional treatment response modeling study used data from a subset of the international Epi25 cohort (Belgium, Finland, Germany). Individuals with epilepsy were included if they had received VPA monotherapy and had available genetic or clinical data. Discovery data (1965-2021, 58% female) were split into a training set (n = 196) and test set (n = 133). Independent validation was performed in a Canadian cohort (2021-2022, n = 156, 40% female). The primary outcome was binary VPA response. Responders achieved ≥12 months of seizure freedom attributed to VPA; nonresponders had >50% seizure recurrence or discontinued VPA due to inefficacy, adverse effects, or unclear reasons. The predictive algorithm integrated features derived from common and rare variants in VPA pharmacokinetic and pharmacodynamic genes, in vitro neuronal VPA response measures, and clinical features. Model performance was assessed using accuracy, predictive values (negative predictive value [NPV]/positive predictive value [PPV]), and area under the curve (AUC). RESULTS: In the independent validation cohort, the multimodal classifier achieved a balanced accuracy of 63% (95% confidence interval [CI] = 52%-73%), NPV of 70% (95% CI = 51%-85%), PPV of 60% (95% CI = 46%-72%), and AUC of .73 (95% CI = .63-.83). Models restricted to single or dual data modalities showed consistently lower predictive performance. SIGNIFICANCE: This proof-of-concept study demonstrates that integrating genetic, cellular, and clinical data enables prediction of VPA treatment response with clinically meaningful accuracy. Although not yet ready for clinical application, this approach supports the feasibility of biomarker-informed ASM selection and may ultimately reduce time to effective seizure control.