Tissue marker KIM-1 flags early kidney cancer recurrence missed by adjuvant criteria

Most patients who undergo surgery for clear-cell renal cell carcinoma never qualify for adjuvant pembrolizumab, the immunotherapy whose use is guided by the KEYNOTE-564 trial criteria. Those criteria rest on tumour stage, grade and nodal status, and the large majority of operated patients fall outside them. A new two-centre exploratory study, published in BMC Cancer, quantifies the consequence: of 117 patients with clear-cell renal cell carcinoma, 23 (19.7%) developed a recurrence within 24 months, and 10 of those 23 recurrences (43.5%) occurred among the 85 patients (72.6%) who did not meet the eligibility criteria.

The ineligible patients whose disease returned were not edge cases that a stricter reading of the rule might have caught. All ten had pT1–2 tumours with pN0 nodal status, ISUP grade 3 or lower and no sarcomatoid differentiation, meaning none triggered any component of the eligibility rule. Early recurrence, in other words, arises substantially among patients whom current criteria classify as low risk and who therefore receive no adjuvant therapy consideration at all under the trial framework.

The researchers asked whether either machine learning or a single immunohistochemical marker, kidney injury molecule-1 (KIM-1), could identify these patients. They applied the KEYNOTE-564 criteria in full, using pT category, ISUP grade, sarcomatoid differentiation and pN status, and trained three models — ridge logistic regression, random forest and gradient boosting — on age, sex, pT, ISUP grade and KIM-1 expression. The models were evaluated by nested cross-validation and by bidirectional centre-to-centre testing of locked models, a demanding design in which a model built at one centre must perform at the other.

Machine learning failed on both counts. No model outperformed the eligibility criteria at matched operating points in any of the three evaluations, and apparent discrimination overstated nested cross-validated performance by up to 0.165 AUC — a cautionary illustration of how optimistic unvalidated performance figures can be. The authors conclude that, in this exploratory cohort, machine learning neither improved on the criteria nor transported between centres.

KIM-1 told a different story, but only within the ineligible group. Among patients outside the criteria, KIM-1 expression was associated with a 24-month recurrence rate of 23.1% versus 6.8%, with a Firth penalised odds ratio of 3.91 (95% CI 1.07–15.48, p = 0.039). Among eligible patients no association was observed (OR 1.63, 95% CI 0.34–7.87, p = 0.53), and the interaction test comparing the two groups was not significant (p = 0.57). A time-to-event sensitivity analysis using a centre-stratified Firth Cox model and including previously excluded patients (n = 121) yielded a corresponding hazard ratio of 3.80 (95% CI 1.15–13.80). These are associations from a small exploratory cohort, not established predictive validity.

The team also tested a composite rule that simply adds KIM-1 positivity to the eligibility criteria. The rule contains no fitted threshold and no estimated parameter, and it was assessed in each centre separately. It captured 19 of the 23 recurrences against 13 for the criteria alone, cutting missed recurrences from ten to four with no recurrence lost: sensitivity rose from 0.565 to 0.826 (McNemar b = 6, c = 0, p = 0.031). The gain came at a cost. Specificity fell from 0.798 to 0.585, meaning more patients without recurrence would be flagged, while negative predictive value rose from 0.882 to 0.932.

The authors are careful about what this trade-off implies. Given the high sensitivity and negative predictive value but reduced specificity, they write that the composite rule warrants prospective evaluation as a tool for stratifying surveillance intensity rather than for selecting adjuvant therapy.

Several limitations frame the findings. The study was exploratory, involved only 117 patients from two centres, and the KIM-1 association was confined to the ineligible subgroup, with a non-significant interaction test. The published version is an early-release, peer-reviewed accepted article that remains subject to further edits before the final version of record. Even so, the central observation stands on its own numbers: a substantial share of early recurrence after clear-cell renal cell carcinoma surgery occurs in patients whom current adjuvant eligibility criteria do not cover, and a simple, parameter-free combination of those criteria with tissue KIM-1 expression detected most of the recurrences the criteria alone miss.

Subject of Research: KIM-1 immunohistochemistry and early recurrence of clear-cell renal cell carcinoma in patients outside adjuvant pembrolizumab eligibility criteria

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

Article References: KAHRAMAN, E. G., ÖZDEMİR, Ö., ÇOBAN, M. M., ARGON, A., HASBAY, E., ÇELİK, S., TAŞKAYNATAN, H., & SELVER, M. A. (2026). KIM-1 immunohistochemistry and early ccRCC recurrence outside adjuvant eligibility criteria: a two-centre exploratory study. BMC Cancer. https://doi.org/10.1186/s12885-026-17191-9

Image Credits: AI Generated

DOI: 10.1186/s12885-026-17191-9

Keywords: clear-cell renal cell carcinoma, KIM-1, adjuvant pembrolizumab, KEYNOTE-564, recurrence, immunohistochemistry, machine learning, nested cross-validation, surveillance, BMC Cancer, early, ccRCC