Comprehensive Summary
Kraus et al focuses on three improvements that can be made to machine learning (ML) models to predict clinical outcomes in psychiatry. These recommendations were made based on past successes of machine learning prediction in other fields such as weather forecasting and drawing parallels to clinical practice. Two requirements that psychiatry machine learning models need to meet for perfect prediction is a thorough understanding of the mechanisms of a complex system and perfect information about the variables present in the system. Weather forecasting ML models have shown an improvement in prediction because of an increase of precise information from new advanced sources such as satellites and balloons. Similar to weather forecasting, Kraus et al recommends obtaining more detailed longitudinal measurements of behavior from individuals. ML models should also use similar criteria to clinicians for diagnosing and treating psychiatric patients such as a psychosis-risk calculator, increasing its dimensionality. While clinicians have access to informative non-verbal behaviors that can’t be incorporated into the model, even with a small fraction of patient information, prediction is still possible. Finally, psychiatric prediction models need a greater sample size of patients for better discrimination of data to predict patient outcomes. All in all, Kraus et al advises ML models need to include personalized detailed data from each individual, greater access to clinician resources, and a greater discriminatory sample size of patients for better prediction of clinical outcomes.
Outcomes and Implications
There are clear benefits to having ML models for precision medicine as they can help create individualized interventions and risk profiles for psychiatric patients. ML models can help identify discriminative biomarkers that distinguish a psychiatric condition from a healthy condition and guide possible treatment options. However, research surrounding ML models in psychiatric contexts have illustrated inadequate prediction of mental disorders without enough information in their data set. Kraus et al outlines various improvements to ML models in psychiatry that will improve its practical application in the clinic. There are a few limitations to take into account such as measuring non-verbal information in the model and addressing ethical concerns about patient burden and confidentiality before ML models can be used effectively. By expanding ML models to clinical practice, diagnosis of psychiatric conditions will be more accurate and efficient and treatment can be more personalized.