TITLE:
A Multimodal Machine Learning Framework for Detecting and Attributing Medication-Induced Cognitive Impairment in Psychiatric Patients
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Medication-Induced Cognitive Impairment, Psychotropic Medications, Neuropsychological Assessment, Deep Learning, Machine Learning, MedBERT, Anticholinergic Burden, Antipsychotics, Bayesian Uncertainty, Drug Attribution, Computational Pharmacopsychiatry, EHR, Cognitive Side Effects
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.5,
May
29,
2026
ABSTRACT: Psychotropic medications are among the most widely prescribed drug classes in the world, and cognitive impairment is among their most consequential and least systematically monitored side effects. A patient stabilized on an antipsychotic who begins to lose processing speed, verbal memory, or executive control does not typically have that change attributed to the medication it gets folded into the illness, missed at brief outpatient reviews, or noticed only when functional decline becomes unmistakable. The cost, measured in lost employment, reduced quality of life, and treatment discontinuation, is substantial. We introduce CogniMed-Net, an end-to-end multimodal machine learning framework that integrates neuropsychological test performance, structured medication profiles, clinical electronic health record (EHR) data, serum biomarkers, and patient demographics to detect, classify, and causally attribute medication-induced cognitive impairment in psychiatric inpatients and outpatients. The framework combines a domain-specific Neuropsychological Encoder, a fine-tuned MedBERT module for medication representation, a Clinical Transformer for EHR feature extraction, a Biomarker MLP, a Longitudinal TCN for trajectory modelling, and a cross-domain attention fusion layer with a Bayesian deep ensemble output providing calibrated uncertainty and causal drug attribution scores. Trained and validated on a retrospective-prospective cohort of 2140 psychiatric patients across five diagnostic categories and twelve psychotropic medication classes, CogniMed-Net achieves 4-class cognitive impairment classification accuracy of 93.4%, AUC-ROC of 0.971, and macro-F1 of 0.907. The Bayesian ensemble achieves Expected Calibration Error (ECE) of 0.024. The model generates model-based drug attribution scores per medication class, reflecting learned associations between pharmacological profiles and impairment patterns, identifying first-generation antipsychotics, benzodiazepines, and tricyclic antidepressants as the highest-attribution drug classes and a continuous cognitive severity score that tracks impairment magnitude independently of diagnostic category. CogniMed-Net establishes a new methodological benchmark for computational pharmacopsychiatry, providing a validated, interpretable architecture for systematic monitoring of psychotropic cognitive side effects in clinical practice.