Evidence for paper review

What the paper claims (and what is actually measured)

The authors develop CLARK, an interpretable longitudinal extension of “latest-value” kidney failure risk equations by adding engineered features derived from repeated lab trajectories (baseline levels, variability, linear/nonlinear trends, changepoints, etc.) on top of demographic factors and last pre-index lab values. They train/test on a retrospective CHS/EHR cohort and model the event as initiation of kidney replacement therapy (KRT) after an index date, censoring at death or loss to follow-up. Methods & outcomes definitions

Evidence strength: discrimination/calibration improvements

In the supplied quantitative results, CLARK’s reported average precision (AP) exceeds the locally refit static baseline CSARK in both horizons. The strongest reported gains occur in the eGFR-only setting: for KRT within 2 years, CLARK AP=0.541 vs CSARK AP=0.516 (ΔAP=+0.024; 95% CI for Δ given as +0.023 to +0.026). For KRT within 5 years, CLARK AP=0.567 vs CSARK AP=0.533 (ΔAP=+0.034; 95% CI +0.033 to +0.035). AP improvements (eGFR-only)

The authors also report that calibration (via Brier score) is modestly improved or comparable when adding longitudinal features, and that c-statistics are higher for CLARK. Calibration & c-statements

Clinical utility: reclassification is horizon- and threshold-dependent

At KDIGO-aligned intervention thresholds, the authors report net reclassification improvement (NRI) that is generally positive for progressors (NRI_KRT>0) but often negative for non-progressors (NRI_No KRT<0), consistent with improved capture of true progressors at the cost of some false positives. They also report that these patterns are more consistent for 5-year thresholds than 2-year thresholds, with minimal/negative effects at the highest 2-year planning threshold (≥40% 2-year risk). KDIGO-aligned NRI patterns

In contrast, under quantile-based strategies (intervene on top k% by predicted risk), they report that gains are largely driven by progressor identification (positive NRI_KRT) with NRI_No KRT “small and not distinguishable from zero,” and provide an example scale: for top 10% interventions in 5-year prediction, an improvement corresponding to +1.8% (≈18 additional KRT cases per 1000 true missed by static model). Quantile-based NRI & example gain

Skeptical reading: what could still limit generalization or causality

• Observational EHR care-pattern bias: trajectory-derived features are built from routinely measured labs; patients with more frequent testing may differ systematically in unmeasured ways (health-seeking behavior, provider intensity). The paper acknowledges EHR-derived measurement patterns can vary and that benefits may differ under different missingness/testing practices. EHR care-pattern limitation
• Outcome ascertainment boundaries: KRT events are defined within CHS records; KRT occurring outside the system could be missed/delayed, potentially biasing event times and training labels. Outcome boundary limitation
• Comparability to clinical equations: the comparison baseline includes published KFRE (via earlier descriptions in the paper) and a locally refit static CSARK to control for cohort/model differences, but bicarbonate scarcity prevented full 6-lab evaluation. So “how much of KFRE-like biology remains” under full lab availability is not tested here. Bicarbonate/6-lab evaluation gap

What would most disprove the headline

A strong disconfirming result would be an external cohort validation showing CLARK (with the same feature engineering and local refits, as appropriate) does not improve discrimination/calibration and does not meaningfully improve progressor-targeting at clinically relevant operating points (KDIGO and capacity-like top-k) compared with the best available latest-value baseline for that external system. External validation necessity stated