Investigators identify the extent to which a machine-learning approach predicts acute improvement for individual depressive symptoms with antidepressants based on pre-treatment symptom scores and EEG measures.
Pranav Rajpurkar, MS
Recent study findings suggest machine learning may be used to identify independent associations of symptoms and electroencephalographic (EEG) features to predict antidepressant-associated improvements in specific symptoms of depression.
Pranav Rajpurkar, MS, and a team of investigators identified the extent to which a machine-learning approach could predict acute improvement for individual depressive symptoms with antidepressants based on pre-treatment symptom scores and EEG measures. The team found their machine-learning algorithm called ElecTreeScore could reliably distinguish patients who responded to treatment from those who did not, based on various symptoms using pre-treatment symptom scores.
Rajpurkar and the team used data collected as part of iSPOT-D, an international, multicenter, randomized, prospective open-label trial being conducted to gather clinically useful predictors and moderators of response to 3 of the most commonly used first-line antidepressants. The study included 1008 adults aged 18-65 years old with a diagnosis of current nonpsychotic major depressive disorder. Participants were unmedicated and randomized in a 1:1:1 ratio to either 8 weeks of treatment with escitalopram (n=162), sertraline (n=176), or extended-release venlafaxine (n=180).
The severity of depressive symptoms was taken at baseline and at the week 8 clinic visit based on the 21-symptom Hamilton Rating Scale for Depression (HRSD-21). Ten of the symptoms were rated on a five-point scale while the other 11 were rated on a three-point scale.
The team continuously recorded electroencephalograms from 26 sites in 5 regions. Resting-state EEG was recorded for 2 minutes while participants were relaxed with their eyes closed and eyes open.
The investigators’ primary objective was to predict improvement in individual symptoms. Improvement was defined as the difference in score for each symptom on the HRSD-21 report from baseline to the week 8 visit using pre-treatment EEG features.
EEG features were generated from pre-treatment EEG recordings at the baseline visit. The team extracted data on the power of the signals in each frequency range at each electrode site.
Rajpurkar and colleagues developed ElecTreeScore, a machine-learning model using gradient-boosted decision trees, to predict improvement in individual symptoms using pre-treatment EEG and baseline HRSD scores. Gradient-boosted decision trees were trained for each of the 21 HRSD categories across several combinations of input features and parameters for the model. Each model was trained on valid combinations of EEG bands, relative and absolute power for frequency bands, electrode site-specific features, and asymmetry features.
The final data set included 518 patients—274 women with a mean age of 39 years old and mean HDRS-21 score improvement of 13. Using five-fold cross-validation, the machine-learning model achieved C index scores of .8 or higher on 12 of 21 depressive symptoms. The highest C index score was .963 (95% CI, .939-1) for loss of insight.
The importance of any EEG feature was higher than 5% for the prediction of 7 symptoms. The most important EEG feature was the absolute delta band power at occipital electrode sites for loss of insight. The use of the EEG and baseline symptom features was linked with a significant increase in the C index for improvement in 4 symptoms: loss of insight (C index increase, .012; 95% CI, .001-.02), energy loss (C index increase, .035; 95% CI, .011=.059), appetite change (C index increase, .017; 95% CI, .003-.03), and psychomotor agitation (C index increase, .02; 95% CI, .008-.032).
The study, “Evaluation of a Machine Learning Model Based on Pretreatment Symptoms and Electroencephalographic Features to Predict Outcomes of Antidepressant Treatment in Adults With Depression,” was published online on JAMA Network Open.