Work

Automated Detection of Schizophrenia Based on Retinal Measures and Neurological Soft Signs: Applications of Interpretable Classification Methods and the Choquet Integral Aggregation Function

Year: 2025

Type: article

Abstract: Background. Schizophrenia is a psychiatric illness with a significant functional impairment. Its diagnosis is based on clinical observation, despite much evidence for its neurobiological basis. More recently, alterations in retinal structure and function were also documented. The retina is a part of the central nervous system, developing embryologically from the same tissue as the brain. Research indicates that retinal morphology measures may be treated as a proxy of brain health. Due to the need to identify schizophrenia biomarkers, we applied a set of automated classifiers encompassing selected retinal and other indirect markers of neuronal system health. Methods. The data came from 59 schizophrenia patients and 61 healthy controls. Interpretable classifiers were also used to differentiate patients with shorter and longer duration of illness. After obtaining the classification results, Choquet integral-based aggregators were adopted to increase discrimination accuracy. Results. Applying Smooth Quadrature-Inspired Choquet aggregators led to a patient vs. control group discrimination accuracy of 93.5%, and accuracy for discrimination between patients with shorter vs. longer duration of illness was 86.25%. In addition, the three-class classification regarding controls and two subgroups of patients reached 88.37% accuracy using 3/8 rule smooth quadrature and triangular norm. Discussion. These results indicate that using measures of retinal morphology and additional variables associated with brain health are useful to classify schizophrenia patients at a level similar to brain imaging-based data. The application of the Choquet integral aggregation function improved the classification results, which is another argument for using this computational method in medical research.

Source: IEEE Access

DOI: 10.1109/access.2025.3579973