Bi-directional AI Framework for Differentiate Psychiatric Disorders and Predicting Symptoms from Drug-Induced Neurotoxicity

Introduction: Diagnosis of mental disorders such as schizophrenia, bipolar disorder, and borderline personality disorder is complicated by the similarity of symptoms, especially in the early stages. The situation becomes even more complicated in the presence of psychoneurological symptoms caused by toxic effects of substances that mimic mental illnesses. There is a need for an intelligent system that can distinguish between these conditions and predict the dynamics of symptoms. Methods: A bidirectional artificial intelligence model was developed, consisting of two modules: a diagnostic classifier (based on XGBoost, LightGBM, CNN) and a prognostic module (based on LSTM/GRU or transformers). Open synthetic and toxicological datasets were used. The model was trained in direct (symptom prediction) and reverse (determination of etiology based on the current state) modes. Efficiency was assessed by classification (accuracy, F1-score, ROC-AUC) and prognostic (MAE, RMSE) metrics. Results: XGBoost demonstrated the highest accuracy (91.2%) in diagnostic classification. The predictive module provided consistently low MAE values ​​when predicting symptoms over a 7- to 30-day horizon. In the inverse analysis mode, the model distinguished endogenous and exogenous symptoms with high probability, especially in cases related to hallucinogens and drug-induced affective lability. Conclusions: The developed AI model demonstrates high accuracy in distinguishing mental disorders from toxic-induced conditions, as well as in predicting symptoms. Its implementation can significantly improve diagnostics and monitoring in psychiatric and toxicological practice, especially with limited clinical information or in outpatient settings.