Using machine learning to improve the discriminative power of the FERD screener in classifying patients with schizophrenia and healthy adults.

Background Facial emotion recognition deficit (FERD) seems to be an obvious feature of patients with schizophrenia and has great potential for classifying patients and non-patients. The FERD screener was previously developed to classify patients from healthy adults. However, an obvious drawback of this screener is that the recommended cut-off scores could enhance either sensitivity or specificity (about 0.92) only, while the other one is at an only acceptable level (about 0.66). Machine learning (ML) algorithms are famous for their feature extraction and data classification abilities, which are promising for improving the discriminative power of screeners. This study aimed to improve the discriminative power of the FERD screener using an ML algorithm. Methods The data were extracted from a previous study. Artificial neural networks were generated to estimate the probability of being a patient with schizophrenia or a healthy adult based on the examinee's responses on the FERD screener (168 items). The performances of the ML-FERD screener were examined using a stratified five-fold cross-validation method. Results Across the five subsets of data, the ML-FERD screener showed extremely high areas under the receiver operating characteristic curve of 0.97-0.99. With the optimized cut-off scores, the average sensitivity and specificity of the ML-FERD screener were 0.90 (0.85-0.93) and 0.93 (0.86-1.00), respectively. Limitations The characteristics of patients were not representative, and the age was mismatched to control group. Conclusion The ML-FERD screener appears to have a better discriminative power to classify patients with schizophrenia and healthy adults than does the FERD screener. [ABSTRACT FROM AUTHOR]