SMASH 2026 · Paper #7 · Result card

Prevalence, Thresholds, and Subgroup Performance Gaps in Clinical Prediction Models Across Five Public Datasets

Zavash Ghorbanpourbaboli, Arina Afshar, Elham Jangravi, Mohammed Ayoub Alaoui Mhamdi, Russell Butler · Bishop's University · Read the paper

With the paper's specification, the Diabetes-130 youngest-minus-oldest sensitivity gap is +0.120 [−0.057, +0.235]; the paper reports +0.104. 19 of 36 alternative analysis choices give a negative gap.

dataset diabetes130attribute agegroups 10-29 vs 60+outcome 30-day readmission (all encounters, patient-level splits)model gradient_boostingthreshold 0.117 (train_prevalence)rows 21,086 teststandardized to π 0.114

Sensitivity gap (10-29 − 60+)

+0.120

95% CI [-0.057, +0.235] — includes 0

PPV gap: raw → standardized

+0.099 → +0.104

gap grows

Prevalence

11.1% vs 11.4%

80 / 2,319 events in test

Gap across analytic choices

-0.24 … +0.15

19 negative / 13 positive of 36 specs

From the abstract

Clinical prediction models are commonly compared across demographic subgroups, but a raw performance gap is not a fixed property of a model. Observed gaps should prompt investigation of metric choice, prevalence, threshold, sample composition, measured case mix, and uncertainty rather than being treated alone as proof of discrimination, fairness, or equivalence.

Evidence

Is the sensitivity gap distinguishable from zero?

Paper vs. your run

Metric	Paper	Your run
Sensitivity difference	+0.104 [-0.070, +0.232]	+0.120 [-0.057, +0.235]	Paper's predictor set, age bands and patient-level resampling; model hyperparameters are ours.
Sensitivity CIs excluding zero (13 comparisons)	7 of 13 (3 sex, 4 age, 0 race)	includes 0	your single comparison
PPV gap after standardization	BRFSS M−F +0.101 → +0.016; CCHS M−F +0.053 → −0.005 (point estimate only; interval spans zero)	+0.099 → +0.104

Try next

Sex gap, diabetes130 (CI excludes zero in paper) — Paper reports −0.049 [−0.084, −0.011]; a CI excluding zero contrasts with the age gap here.
```
{"dataset": "diabetes130", "attribute": "sex"}
```
Race gap, diabetes130 (AfricanAmerican vs Caucasian) — Paper reports −0.040 [−0.090, +0.006]; tests whether a race comparison also shows multiverse instability.
```
{"dataset": "diabetes130", "attribute": "race"}
```
Age gap, CDC Diabetes (large robust gap in paper) — Paper reports −0.207 [−0.241, −0.170]; a CI that excludes zero contrasts with the fragile diabetes130 age gap.
```
{"dataset": "cdc_diabetes", "attribute": "age"}
```

Read this before sharing

Model hyperparameters and variable encodings are this runbook's; the predictor set and age bands follow the paper.
No fairness verdict is implied.