The risk of round numbers and sharp thresholds in clinical practice (www.nature.com)

Researchers developed an interpretable "glass-box" machine learning framework to expose how round-number clinical thresholds (e.g., creatinine 3.5 mg/dL, BUN ~40 mg/dL, systolic BP <80 mmHg) create sharp discontinuities and counter‑causal non‑monotonicities in observed mortality risk. Using GAMs (boosted-tree components) alongside simulation experiments and longitudinal analyses of real hospital datasets (MCHD, MIMIC-IV), they show that threshold-guided treatment decisions can produce surprising population-level patterns — for example, very high serum creatinine (>5 mg/dL) appearing associated with lower mortality because aggressive treatment is concentrated above a threshold. The team formalized two statistical tests to automatically detect discontinuities and paradoxical risk shapes in feature-specific risk functions. For the AI/ML community, the work is a cautionary technical message: predictive models trained on observational clinical data can learn artifacts of treatment policy rather than underlying biology, leading to dangerous recommendations (e.g., de‑prioritizing patients who previously received effective interventions). The paper demonstrates how combining interpretable GAMs, simulation-based causal intuition (varying treatment benefit shapes and adherence), and automated shape-tests can detect these artifacts, arguing for models that explicitly account for treatment assignment, continuous reassessment of protocol thresholds, and longitudinal validation to ensure safe, clinically aligned AI deployment.