Age-group transportability of explainable machine-learning models for cancer-specific survival in early-onset colorectal cancer: a temporally validated U.S. SEER study

Preprint — not peer reviewed. This study evaluates the transportability and recalibration of colorectal cancer survival models across age groups using U.S. SEER registry data. Abstract Background: Early-onset colorectal cancer (EO-CRC, diagnosis at ages 18–49) is rising in incidence, yet prognostic models are predominantly developed in average-onset (AO) populations. Whether AO-trained survival models transport to EO patients without refitting — and whether simple recalibration repairs any degradation — has not been tested under a preregistered, temporally validated design. Methods: Population-based cohort study using SEER 17-registry data (first malignant primary colorectal adenocarcinoma, diagnosed 2010–2019 at ages 18–84; N = 216,975). Models were trained on 2010–2016 diagnoses and tested temporally on 2017–2019 diagnoses (EO: 20,781 train / 9,551 test; AO: 130,728 / 55,915). The primary outcome was cancer-specific survival; the primary horizon (60 months) was locked by a prespecified follow-up-adequacy gate before any model was fitted. AO-trained Cox, XGBoost-Cox, and fixed-horizon XGBoost models were applied frozen — preprocessing included — to EO test patients. Prespecified materiality criteria defined transport degradation: paired ΔC-index ≥ 0.02 with 95% CI excluding 0; calibration slope outside 0.80–1.20; or |calibration-in-the-large (CITL)| > 0.20. Four recalibration variants, fitted on EO training data only, were evaluated against a prespecified adequacy range (slope 0.90–1.10 and |CITL| ≤ 0.10). SHAP analyses were descriptive only. Results: Discrimination transported essentially without loss: paired ΔC +0.0001 (95% CI −0.0014 to +0.0014) for Cox and −0.0014 (−0.0027 to −0.0003) for XGBoost-Cox; neither met the prespecified materiality criterion. The transported fixed-horizon model met the miscalibration criterion (CITL +0.413; slope 0.902, in range), so the disjunctive transportability hypothesis was supported on the calibration limb only. However, every fixed-horizon model under-predicted 5-year cancer-specific death in the temporal test window, including models evaluated in their own age group (EO-on-EO CITL +0.300; AO-on-AO +0.384); transport therefore cannot be identified as the cause. No recalibration variant met the prespecified adequacy range (slope was repaired to 0.945–1.081, but CITL remained +0.23 to +0.35). Among EO-trained models, the full clinicopathologic Cox model outperformed a stage-only model (paired ΔC −0.0556), while machine-learning gains over the full Cox benchmark were small (XGBoost-Cox +0.0078; random survival forest +0.0022). In prespecified sensitivity analyses, discrimination findings were unchanged; the transported model’s miscalibration criterion remained met after excluding rectal primaries (CITL +0.294) but was no longer met after excluding 2019 diagnoses (CITL −0.008; test window 2017–2018), where one prespecified recalibration variant also achieved adequacy. Conclusions: In a preregistered temporal validation, relative risk ordering learned in AO-CRC transferred to EO-CRC patients with negligible discrimination loss, but absolute risk calibration failed prespecified criteria across models and age groups in the temporal test window, and prespecified EO-only recalibration did not repair it there — although when the test window was restricted to 2017–2018, the transported model was essentially calibrated in the large and one prespecified recalibration variant achieved adequacy. Risk-score transport appears viable; absolute-risk deployment requires temporally current calibration. Keywords: early-onset colorectal cancer; transportability; temporal validation; machine learning; survival analysis; calibration; SEER