Pharmacoepidemiology and drug safety · 2026
Unsupervised Ensemble Learning for Active Drug-Induced Liver Injury Surveillance: Integrating Pharmacokinetic Burden With Enzyme Trajectories.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Nakkaratniyom Thawatchai, Hirunrassamee Sanita, Boonchoo Thapana, Sooksriwong Cha-Oncin
Аннотация
PURPOSE: Traditional drug-induced liver injury (DILI) surveillance relying on static laboratory thresholds frequently misses early kinetic evolution. To address this, a fundamentally drug-agnostic unsupervised ensemble framework, integrating pharmacokinetic (WCEDR) and polypharmacy (HAMPIS) features, was developed for early DILI detection. Antiepileptic drugs (AEDs) were utilised as a proof-of-concept validation cohort.
METHODS: Using Thailand's national electronic health record database (2021-2024), a multi-view ensemble (Isolation Forest, Local Outlier Factor, One-Class Support Vector Machine) evaluated longitudinal physiological deviations. Feature engineering prioritised acute kinetic volatility to mitigate irregular sampling intervals. Diagnostic performance was benchmarked against standard criteria (ALT > 3 × ULN) and validated through blinded expert adjudication of 140 clinical episodes.
RESULTS: From 616 425 treatment episodes, the framework isolated 86 high-probability anomalies. Benchmarking revealed the model captured 80.2% of rule-based positives while proactively identifying 17 AI-only sub-threshold episodes (19.8%) exhibiting distinct enzymatic velocity; over half (52.9%) were clinically confirmed as DILI. Blinded adjudication yielded a clinical plausibility rate of 69.8% and a negative predictive value of 92.6%. Operationally, the framework achieved a number needed to review (NNR) of 1.43, substantially optimising triage efficiency. Digital phenotyping stratified alerts into three distinct archetypes: complex polypharmacy, atypical pharmacokinetic burden, and active idiosyncratic injury.
CONCLUSION: By prioritising kinetic volatility and multidimensional pharmacological context, the unsupervised ensemble functions as a high-precision, complementary surveillance tool. It successfully detects early-onset idiosyncratic reactions frequently overlooked by rigid static thresholds, offering a scalable blueprint for proactive pharmacovigilance.
METHODS: Using Thailand's national electronic health record database (2021-2024), a multi-view ensemble (Isolation Forest, Local Outlier Factor, One-Class Support Vector Machine) evaluated longitudinal physiological deviations. Feature engineering prioritised acute kinetic volatility to mitigate irregular sampling intervals. Diagnostic performance was benchmarked against standard criteria (ALT > 3 × ULN) and validated through blinded expert adjudication of 140 clinical episodes.
RESULTS: From 616 425 treatment episodes, the framework isolated 86 high-probability anomalies. Benchmarking revealed the model captured 80.2% of rule-based positives while proactively identifying 17 AI-only sub-threshold episodes (19.8%) exhibiting distinct enzymatic velocity; over half (52.9%) were clinically confirmed as DILI. Blinded adjudication yielded a clinical plausibility rate of 69.8% and a negative predictive value of 92.6%. Operationally, the framework achieved a number needed to review (NNR) of 1.43, substantially optimising triage efficiency. Digital phenotyping stratified alerts into three distinct archetypes: complex polypharmacy, atypical pharmacokinetic burden, and active idiosyncratic injury.
CONCLUSION: By prioritising kinetic volatility and multidimensional pharmacological context, the unsupervised ensemble functions as a high-precision, complementary surveillance tool. It successfully detects early-onset idiosyncratic reactions frequently overlooked by rigid static thresholds, offering a scalable blueprint for proactive pharmacovigilance.