MentalAId: an improved DenseNet model to assist scalable psychosis assessment.

BACKGROUND

The escalating mental health crisis during and post-COVID-19 underscores the urgent need for scalable, timely, cost-effective assessment solutions for general psychotic disorders. Regretfully, traditional symptom-based, one-to-one assessment face inherent limitations in large-scale and longitudinal screening, likely delaying early intervention.

METHODS

We developed MentalAId, an improved densely connected convolutional network (DenseNet) model, to assist automated psychosis recognition, leveraging accessible routine laboratory data without requiring additional specialized tests. MentalAId learned subtle variations in 49 routine clinical hematological tests and two demographic variables (sex and age) across 28,746 individuals spanning four distinct cohorts: psychotic inpatients (n = 9,271), non-psychotic inpatients with various diseases (n = 14,508), healthy controls (n = 1,826), and drug-naïve first-episode psychosis (FEP) patients (n = 3,141).

RESULTS

The MentalAId model achieved high accuracy in generally discriminating psychoses from both healthy individuals and patients with other physical diseases, achieving 93.3% accuracy and AUC of 0.983. Further validating its robustness, MentalAId demonstrated high performance under real-world clinical conditions, accurately handling extreme values and missing values, with accuracies of 92.4% and 92.0%, respectively. Even encompassing the drug-naïve FEPs, MentalAId maintained an accuracy of 91.9%, underscoring its translational potential for early FEP recognition. Interpretability analyses identified indirect bilirubin, direct bilirubin, and basophil ratio as potential metabolic indicators.

CONCLUSIONS

Collectively, MentalAId offers an accessible, affordable, and scalable solution to assist timely psychosis assessment and monitoring, irrespective of the complexity of pathology or manifestation of symptoms. By requiring standard blood tests solely, it can be easily integrated into existing healthcare workflow. This empowers long-term and population-wide monitoring of disease progression and prognosis, particularly during public health crises like COVID-19.