Climbing the ladder: a ranking approach to burnout prediction

Validated psychological assessment tools, such as the Shirom-Melamed Burnout Measure (SMBM), are essential for reliably assessing burnout. However, their reliance on active, self-reported input limits their suitability for continuous monitoring and early detection, and introduces the potential for human bias. The SMBM specifically targets the energy depletion component of burnout, with items organized into three subscales: Physical Fatigue (PF), Cognitive Weariness (CW), and Emotional Exhaustion (EE). In the present work, we investigate the feasibility of predicting burnout risk unobtrusively using the preceding trajectory of passive physiological data from wearable devices, supplemented by baseline demographic and occupational information. We evaluate classification, regression, and learning-to-rank formulations for the prediction of SMBM subscale scores on a 9-month real-world dataset of 239 workers, using both aggregate-based and sequential models. Binary classification yields modest performance [ROC AUC: PF (0.66), CW (0.67), EE(0.56)], and regression models offer negligible gains over naïve benchmarks. However, rank-based metrics suggest relative burnout severity can be partially inferred from wearable signals. Motivated by this, we propose a siamese recurrent neural network, explicitly tailored for sequential wearable data and optimized for pairwise risk estimation. Results show improved alignment with the ordinal nature of burnout scores for PF (Spearman's <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>ρ</mml:mi> <mml:mo>=</mml:mo> <mml:mn>0.29</mml:mn></mml:math> , Normalized Discounted Cumulative Gain <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mo>=</mml:mo> <mml:mn>0.93</mml:mn></mml:math> ) and CW (Spearman's <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>ρ</mml:mi> <mml:mo>=</mml:mo> <mml:mn>0.25</mml:mn></mml:math> , Normalized Discounted Cumulative Gain <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mo>=</mml:mo></mml:math> 0.89), whereas assessing EE may require additional modalities. Although real-world implementation remains challenging, ranking-based techniques could pave the way for more effective burnout risk monitoring.