Models that used coded clinical data from electronic health records predicted antidepressant treatment response in general but not response to specific antidepressants in patients with major depressive disorder, according to a study published online in JAMA Network Open.
“In this analysis of electronic health records from more than 81,000 individuals across two health systems, we identified machine learning models that predicted achievement of treatment stability, a proxy for effectiveness, based solely on coded clinical data already available instead of incorporating research measures or questionnaires,” researchers wrote.
Among the 81,630 adults in the analysis with major depressive disorder, 55,303 continued the same prescription for at least 90 days, which researchers considered a stable treatment response.
Machine learning models that predicted antidepressant treatment stability had areas under the receiver operating characteristic curve that ranged from 0.60 to 0.66, suggesting modest discrimination. However, building models with treatment-specific predictors instead of general predictors did not improve prediction, researchers reported.
“This may reflect the observation that much of antidepressant response may be considered to be placebo-like or nonspecific,” they wrote. “That is, although antidepressants consistently demonstrate superiority to placebo, placebo response is substantial such that nonspecific predictors may outperform drug-specific ones.”
That is not to suggest that medication-specific predictors do not exist, researchers clarified, but that identifying them will require other strategies.
“Although greater discrimination is likely required for clinical application” of study findings, researchers concluded, “the results provide a transparent baseline for such studies.”
Hughes MC, Pradier MF, Ross AS, McCoy TH Jr, Perlis RH, Doshi-Velez F. Assessment of a prediction model for antidepressant treatment stability using supervised topic models. JAMA Network Open. 2020;3(5):e205308.