From data to diagnosis: A logical learning method to enhance interpretability in bipolar and major depressive disorder identification

The significance of intelligent diagnosis technology in enhancing early detection efficiency is paramount. However, the complexity of machine learning algorithms often hampers result interpretability. This paper proposes an interpretable diagnostic method named logical learning, which combines multi-attribute value theory, machine learning, and optimization techniques. It simulates physicians’ diagnostic rules/logic using an interactive value function, considering the marginal values and importance of features, along with their interactions. A variant of a gradient descent optimization algorithm and cross-validation are utilized to estimate a comprehensive decision model from historical diagnosis data. The logical learning method is applied to distinguish bipolar disorder (BD) and major depressive disorder (MDD) using the electronic medical records of 6157 patients from a large hospital in western China. It provides the degree of contribution of each feature to the diagnosis and explicitly indicates which symptoms’ presence, abnormally high or low biomarkers have significant contributions to the diagnosis of BD or MDD. With an AUC (area under the curve) of 0.851 and an accuracy of 0.803, the proposed method demonstrates superior performance than traditional machine learning.