Detecting individuals with severe mental illness using artificial intelligence applied to magnetic resonance imaging

Abstract

Background: Identifying individuals at risk for severe mental illness (SMI) is crucial for prevention and early intervention strategies. While MRI shows potential for case identification even before illness onset, no practical model for mental health risk monitoring has been developed. This study aims to develop an initial version of an efficient and practical model for mental health screening among at-risk populations.

Methods: A deep learning model known as Multiple Instance Learning (MIL) was adopted to train and test a SMI detection model with clinical MRI scans of 14,915 patients with SMI (age 32.98 ± 12.01 years, 9102 women) and 4538 healthy controls (age 40.60 ± 10.95 years, 2424 women) in the primary dataset. Validation analysis was conducted in an independent dataset with 290 patients (age 28.08 ± 10.95 years, 169 women) and 310 healthy participants (age 33.55 ± 11.09 years, 165 women). Another three machine learning models of ResNet, DenseNet and EfficientNet were used for comparison. We also recruited 148 individuals receiving high-stress medical school education to characterize the potential real-world utility of the MIL model in detecting risk of mental illness.

Findings: Similar performance of successful differentiation of individuals with SMI and healthy controls was observed for the MIL model (AUC: 0.82) and other models (ResNet, DenseNet, EfficientNet, 0.83, 0.81, and 0.80 respectively). MIL had better generalization in the validation test than other models (AUC: 0.82 vs 0.59, 0.66 and 0.59), and less drop-off in performance from 3.0T to 1.5T scanners. The MIL model did better in predicting clinician ratings of distress than self-ratings with questionnaires (84% vs 22%) in the medical student sample. Brain regions that contributed to SMI identification were mainly neocortical, including right precuneus, bilateral temporal regions, left precentral/postcentral gyrus, bilateral medial prefrontal cortex and right cerebellum.

Interpretation: Our digital model based on brief clinical MRI protocols identified individual SMI patients with good accuracy and high sensitivity, suggesting that with incremental improvements the approach may offer potentially useful aid for early identification and intervention to prevent illness onset in vulnerable at-risk populations.

Funding: This study was supported by the National Natural Science Foundation of China, National Key Technologies R&D Program of China, and Sichuan Science and Technology Program.

Keywords: Individual diagnosis; Multiple instance learning; Neuroimaging; Screening; Severe mental illness.

Fig. 1

The framework of the proposed method. Abbreviations: FC, fully connected layer; FE, Feature extraction; GAP, global average pooling; HC, healthy controls; SMI, severe mental illness.

Fig. 2

The boxplots of diagnostic metrics about Area Under the Curve (AUC), Accuracy, Sensitivity and Specificity of Residual Neural Network, Dense Convolutional Network, Efficient Neural Network, and Multiple Instance Learning with and without class activate map module in primary dataset (Dataset 1), validation dataset (Dataset 2) and real-world testing dataset (Dataset 3). Two-tailed adjusted p values are obtained and represented with asterisk, with ∗ indicating p < 0.05, ∗∗ indicating p < 0.01, and ∗∗∗ indicating p < 0.001. Abbreviations: AUC, Area Under the Curve; MIL, Multiple Instance Learning; ResNet, Residual Neural Network; DenseNet, Dense Convolutional Network; EfficientNet, Efficient Neural Network; w/o-cam, without class activate map module.

Fig. 3

The class activate map of the patient/control differentiation during the decision process of Multiple Instance Learning. Warmer areas on the attention map indicated regions with relative higher contribution weight in patient or control identification. Upper panel: average attention map of brain regions that contributed to the patient identification. Lower panel: average attention map of brain regions that contributed to healthy control identification. Abbreviations: L, left; R, right.