Differentiation between psychotic and non-psychotic major depression by the tabular prior-data fitted network.

BACKGROUND: Misdiagnosing psychotic major depression (PMD) as non-psychotic major depression (NPMD) can lead to poor treatment outcomes. This study aims to develop and validate a machine learning-based model using electronic medical record (EMR) data and the Tabular Prior-data Fitted Network (TabPFN) model to distinguish between PMD and NPMD. METHODS: A total of 666 patients with PMD and 808 patients with NPMD from January 2020 to February 2025 were included. The Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was first applied for feature screening, and the TabPFN model was constructed based on the screening results. The TabPFN model was compared with seven traditional ML models, and SHapley Additive exPlanations (SHAP) were used to interpret the predictions. RESULTS: The TabPFN model demonstrated robust performance compared to other traditional ML models, achieving an area under the curve (AUC) of 0.798. According to SHAP values, thyroxine (T4) was among the most important features along with age. In particular, elevated T4 levels and younger age were associated with a higher likelihood of PMD. LIMITATIONS: The use of EMR data from a single center may limit the generalizability of our findings to other clinical settings or populations. CONCLUSIONS: In comparison to traditional ML models, the TabPFN model constructed in our study based on eight important variables that were selected via LASSO has promising potential for aiding in the diagnosis of PMD and NPMD. Based on its special algorithmic advantages and simple preprocessing workflow, this model can provide insights into developing auxiliary diagnostic tools to assist in identifying psychiatric disorders.