Machine Learning Predicts Progression to Schizophrenia Within 5 Years
Machine learning models used routine clinical data from electronic health records to predict progression to schizophrenia and bipolar disorder within 5 years for patients with pre-existing mental illness, according to a study published in JAMA Psychiatry.
“The model predicting schizophrenia performed substantially better than the model predicting bipolar disorder, likely due to heterogenic clinical manifestations of the latter,” wrote corresponding author Lasse Hansen, MSc, PhD, of Aarhus University, Aarhus, Denmark, and study coauthors.
The study included 24,449 patients aged 15 to 60 years with a total of 398,922 outpatient contacts for psychiatric services in Denmark between 2013 and 2016. Researchers investigated whether machine learning models trained on clinical data could predict diagnostic progression to schizophrenia and bipolar disorder. They gauged performance of the models they developed using area under the receiver operating characteristic curve (AUROC).
The best model for predicting transition to the occurrence of either schizophrenia or bipolar disorder had an AUROC of 0.70 on a training set and 0.64 on a test set, according to the study. The model had a sensitivity of 9.3%, a specificity of 96.3%, and a positive predictive value (PPV) of 13.0% at a 4% predicted positive rate threshold.
>>RELATED: Researchers Advise Screening for Manic Symptoms in Patients With Schizophrenia
A model predicting schizophrenia alone demonstrated the best performance, with an AUROC on the test set of 0.80. At a 4% predicted positive rate, the model had a sensitivity of 19.4%, a specificity of 96.3%, and a PPV of 10.8%, the study found.
“These findings suggest that detecting progression to schizophrenia through machine learning based on routine clinical data is feasible,” researchers wrote, “which may reduce diagnostic delay and duration of untreated illness.”
Bipolar disorder was more difficult to predict, however, with the best model achieving an AUROC of 0.62, a sensitivity of 9.9%, a specificity of 96.2%, and a PPV of 8.4% at a predicted positive rate of 4%.
Text-based predictors from clinical notes proved particularly informative for the models, researchers reported. The most important words were related to hospital admission or psychiatric symptoms.
“Indeed,” the authors wrote, “models trained with both structured and text-based predictors performed practically equivalent to models trained with only text-based features; this underscores the importance of text in clinical prediction modeling within psychiatry.”
Reference
