Auditing Construct Overlap in Explainable Machine Learning: Evidence from Burnout-Depression Prediction Across Student Cohorts

Apparent robustness can mask fundamental problems: The study found that an explainable AI model showed seemingly impressive and stable results across multiple student populations (Kendall τ=1.0 for top predictors), but these results were largely artifacts of how the outcome was constructed rather than genuine predictive insights.

Overlapping definitions undermine interpretability: When researchers used a statistical technique (residualization) to remove shared variance between correlated variables, the model's performance collapsed dramatically (R² dropped from 0.41 to 0.016 for burnout prediction), revealing that the model was exploiting definitional overlap rather than discovering meaningful relationships.

High prediction accuracy doesn't guarantee real-world usefulness: Despite showing R² values of 0.41-0.49 across datasets, the model's prediction intervals were too wide (averaging 35.4 units on a 0-100 scale) to reliably identify individual students at risk, making it unsuitable for actual deployment in mental health screening.

Explainability requires construct validation: The paper demonstrates that explainable AI systems need systematic auditing to verify that identified important features reflect genuine causal relationships rather than statistical artifacts created by how outcome variables were measured or defined.

A practical audit protocol is now available: The authors provide a transferable residualization method that any study combining correlated predictors and outcomes should apply before claiming to have identified meaningful feature importance rankings.