Predicting conversion to psychosis using machine learning: response to Cannon.

Area of Science:

• Neuroscience
• Computational Psychiatry

Background:

• Machine learning models demonstrated high accuracy (up to 90%) in predicting psychosis conversion in individuals at clinical high risk (CHR) using the North American Prodrome Longitudinal Study-3 (NAPLS-3) dataset.
• Testing model generalization on an independent dataset is crucial for validating predictive capabilities.

Purpose of the Study:

• To evaluate the generalizability of a machine learning model trained on the NAPLS-3 dataset by testing it on the independent NAPLS-2 dataset.
• To assess the performance of identical machine learning algorithms on a previous iteration of the study data.

Main Methods:

• Standard machine learning algorithms were employed.
• Models were trained on the NAPLS-3 dataset to predict psychosis conversion.
• The trained models were then tested on the NAPLS-2 dataset.

Main Results:

• Significant differences were observed in features between NAPLS-2 and NAPLS-3 participants.
• All machine learning models performed above chance, with Naive Bayes and random forest showing the best results.
• Model performance on the NAPLS-2 dataset did not replicate the high accuracy seen with the NAPLS-3 dataset alone.

Conclusions:

• Machine learning models trained on one dataset can generalize to independent datasets, but performance may be reduced.
• Current model performance is insufficient for direct clinical application in psychosis prediction.
• Discrepancies in participant features between datasets likely contributed to the performance limitations.