Predicting the need for electroconvulsive therapy via machine learning trained on electronic health record data.

Area of Science:

• Psychiatry and Mental Health
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Electroconvulsive therapy (ECT) is a vital treatment for severe mental illness.
• Delays in ECT initiation can negatively impact patient outcomes.
• Predicting the need for ECT can facilitate prompt treatment.

Purpose of the Study:

• To develop and validate a machine learning model to predict the need for ECT in psychiatric inpatients.
• To utilize electronic health record (EHR) data for predicting ECT initiation.

Main Methods:

• Utilized EHR data from 41,610 adult patients across 164,961 admissions (2013-2021).
• Trained extreme gradient boosting models on the 7th day of inpatient stay using an 85% training set.
• Validated model performance on a 15% test set, defining ECT initiation as occurring >7 and ≤67 days post-admission.

Main Results:

• The model achieved an Area Under the Curve (AUC) of 0.94 in the test set.
• Demonstrated high specificity (98%) and negative predictive value (99%) for predicting ECT need.
• Key predictors included highest suicide assessment score and mean Brøset violence checklist score.

Conclusions:

• Electronic health record data can effectively predict the need for electroconvulsive therapy.
• Predictive modeling holds potential for optimizing treatment timelines and improving patient care in psychiatry.