UTILIZING AI MODELS TO IDENTIFY AND PREDICT MOOD TRANSITION PATTERNS IN PATIENTS WITH BIPOLAR DISORDER

Abstract Bipolar disorder is characterized by recurrent episodes of depression and mania, making diagnosis and treatment challenging, often leading to poor prognosis. Predicting future episodes based on the assessment of current and past emotional phases in patients can offer valuable insights for clinicians, aiding in the prevention of misdiagnosis and relapse, and in intervention strategies. This study included 812 patients with bipolar disorder from 9 centers, who were assessed online or offline using BDCC, YMRS, and HAMD tools. We employed four AI methodologies to learn from the assessment results and time series, ultimately predicting future emotional phases of the patients. By learning from current assessment results to predict the next emotional phase, the AUCs of the four models were 0.81, 0.70, 0.84, and 0.81, respectively. External validation on 30-day and 90-day emotional phase follow-ups of patients showed prediction accuracies of 0.9, 0.65, 0.86, and 0.62 for each model. While predictive efficacy decreased over time, the average AUC remained above 0.75 at 180 days. Our findings suggest that utilizing big data and AI methods can effectively learn the characteristics of emotional phase transitions in bipolar patients, achieving high accuracy. This model holds significant value for clinical diagnosis and medication guidance for clinicians