KIM-1 immunohistochemistry and early ccRCC recurrence outside adjuvant eligibility criteria: a two-centre exploratory study

Background Adjuvant pembrolizumab is offered to patients whose tumours meet the KEYNOTE-564 eligibility criteria. These criteria are based on stage, grade and nodal status, and the large majority of operated patients do not meet them and are therefore not candidates for adjuvant pembrolizumab under the trial criteria. We quantified how much early recurrence arises among ineligible patients and asked, in an exploratory analysis, whether machine learning or a single immunohistochemical marker, kidney injury molecule-1 (KIM-1), can identify the patients concerned. Methods We studied 117 patients with clear-cell renal cell carcinoma operated at two centres. The endpoint was recurrence within 24 months. KEYNOTE-564 criteria were applied in full using pT category, ISUP grade, sarcomatoid differentiation and pN status. Ridge logistic regression, random forest and gradient boosting were trained on age, sex, pT, ISUP and KIM-1, and evaluated by nested cross-validation and by bidirectional centre-to-centre testing of locked models. A composite rule combining eligibility with KIM-1 positivity — containing no fitted threshold and no estimated parameter — was assessed in each centre separately. Effect estimates used Firth penalised logistic regression with profile-likelihood intervals; a centre-stratified Firth Cox model was added as a time-to-event sensitivity analysis. Results Twenty-three patients (19.7%) recurred within 24 months. Eighty-five patients (72.6%) did not meet the criteria, and 10 of the 23 recurrences (43.5%) occurred among them. All ten were pT1–2 with pN0 disease, ISUP grade ≤ 3 and no sarcomatoid differentiation, so none triggered any component of the rule. Within this ineligible group, KIM-1 expression was associated with a 24-month recurrence rate of 23.1% versus 6.8% (Firth OR 3.91, 95% CI 1.07–15.48, p = 0.039), whereas no association was observed among eligible patients (OR 1.63, 95% CI 0.34–7.87, p = 0.53); the interaction test was not significant ( p = 0.57). In a time-to-event sensitivity analysis including previously excluded patients ( n = 121), the corresponding hazard ratio was 3.80 (95% CI 1.15–13.80). No machine-learning model outperformed the criteria at matched operating points in any of three evaluations; apparent discrimination overstated nested cross-validated performance by up to 0.165 AUC. The composite rule captured 19 of 23 recurrences against 13 for the criteria alone, reducing missed recurrences from ten to four with no recurrence lost (sensitivity 0.826 versus 0.565; McNemar b = 6, c = 0, p = 0.031); specificity fell from 0.798 to 0.585 and negative predictive value rose from 0.882 to 0.932. Conclusions A substantial share of early recurrence occurs in patients who do not meet current adjuvant eligibility criteria. In this exploratory cohort, machine learning did not improve on the criteria or transport between centres, whereas tissue KIM-1 was associated with recurrence among ineligible patients. Given high sensitivity and negative predictive value but reduced specificity, the composite rule warrants prospective evaluation as a tool for stratifying surveillance intensity rather than for selecting adjuvant therapy.