Using Gait/balance To Identify Healthy Individuals Feeling Depressed: An Exploratory Study Using Machine Learning

PURPOSE: To identify healthy individuals who report feeling depressed at the moment (without a current diagnosis of major depressive disorder (MDD)) using machine learning (ML) on balance and self-selected walking gait data. METHODS: Participants (n = 100, depressed = 20) between the ages of 18-36 (24.15 ± 3.95) were asked to complete the Profile of Moods Survey-Short Form (POMS-SF). Using the depression scores from that scale, participants were classified as depressed (D) or not depressed (ND). After completing the POMS-SF, participants completed the modified Clinical Test of Sensory Interaction in Balance (mCTSIB) and a 2-minute walk around two cones placed 6 m apart using seven APDM mobility monitors. After feature selection, multiple classifier ML models were used to identify individuals as D/ND using a) gait, b) all four conditions of the mCTSIB (eyes open/closed, firm/foam surface) combined and separately, and c) a combination of gait and all four balance conditions. Due to the imbalanced dataset, F1-means were calculated for harmonic accuracy means and precision recall (PR) was used to measure performance. Validation was completed using a Leave One Out method and models with the best PR and F1 scores were selected. RESULTS: There were no significant differences in age/height/weight/sex/BMI between the two groups. The most accurate model for identifying feelings of depression was a decision tree classifier (DTC) (F1 = 0.84, PR = 0.84), followed by a DTC using combined gait and all balance conditions (F1 = 0.81, PR = 0.81). All balance conditions combined (MLP) and individually had the same accuracy (F1 = 0.80, PR = 0.80). CONCLUSION: Our findings offer promising results: gait and balance may be an accurate and precise way to identify healthy individuals who report feeling depressed. Although the 2-minute walk was the most accurate and precise model, capturing this measure in the real world may be impractical. However, our findings suggest that a 30-second balance test with eyes open and feet on the ground (Condition 1 of the mCTSIB) may provide comparable results. Therefore, future researchers should seek to replicate our findings by utilizing methods, such as computer vision or single sensors and balance to create more robust models in identifying healthy individuals who report feeling depressed.