Clear cell renal cell carcinoma (ccRCC) frequently returns after surgery, and routine clinicopathological tools can struggle to sort patients reliably into risk groups that guide adjuvant therapy and follow-up. A multimodal predictive recurrence score (MPRS) was developed using information already generated in standard care: clinical features, contrast-enhanced CT and histopathological whole-slide images. Built across six centres with additional histopathology from a public resource, the approach achieved strong discrimination and addressed important misclassifications made by widely used clinicopathological scores. By drawing signal from several data streams rather than one, the model offers a more practical route to consistent risk assessment and may help clinicians target adjuvant therapy where it is more likely to benefit patients while avoiding overtreatment.
Built From Routine Data Across Multiple Centres
The work assembled 1,648 patients with ccRCC, including 1,145 people with stage I–III disease from six Chinese centres and additional histopathology from a public database to strengthen training for the pathology branch. Within the in-house set, 654 patients formed the training cohort, 134 made up the internal validation cohort and 357 were reserved for external validation. To build the integrated model, 550 patients in the training cohort had all three modalities available. In total, 4,001 whole-slide images and 1,098 contrast-enhanced CT scans were processed, with deep learning and radiomic or pathomic features extracted before fusion into the MPRS. A threshold derived from time-dependent analysis was used to split patients into low- and high-risk groups for disease-free survival.
Must Read: Improved MRI Scoring for Clear Cell Renal Cancer
Baseline characteristics reflected typical surgical ccRCC populations. Most patients were male and the median age was 60 years. Stage I disease and histopathological grade 2 predominated, while necrosis and sarcomatoid differentiation were less common. Over a median follow-up of 49 months, just over one in six patients experienced recurrence. Where only one or two data types were available, unimodal or bimodal models were trained, but the central focus remained the fully integrated score. The design emphasised practicality by leveraging data produced during routine diagnostic and treatment pathways, avoiding the need for costly molecular profiling.
Stronger Stratification Than Established Tools
Across training, internal validation and external validation cohorts, the multimodal score delivered higher discrimination than either unimodal models or familiar clinical tools, including the Leibovich score, the UISS score and the KEYNOTE-564 risk grouping. Performance in independent cohorts remained strong, with C-index values of 0.886 in internal validation and 0.838 in external validation. When the two validation cohorts were combined, classification of low- and high-risk groups showed a balanced profile, with sensitivity of 85.4% and specificity of 78.2%. This contrasted with imbalances observed for clinicopathological tools, where high specificity often came at the cost of low sensitivity or vice versa.
Correcting clinically relevant misclassifications was a notable feature. Among patients who recurred but were labelled low risk by KEYNOTE-564, the multimodal score reclassified 83.3% as high risk, addressing the risk of undertreating those individuals. Conversely, among non-recurrent patients mislabelled as intermediate or high risk by KEYNOTE-564, 57.7% were reclassified as low risk, reducing the likelihood of unnecessary adjuvant therapy. Similar trends were seen when compared with the Leibovich and UISS scores, where a large share of their respective false negatives and false positives were reassigned correctly. Multivariable analyses confirmed that high-risk assignment by the multimodal score was independently associated with worse disease-free survival after accounting for basic clinical factors.
Interpretability, Error Patterns and Limits
Interpretability analyses offered reassurance that the model’s focus aligns with known prognostic features. Clinical variable importance placed histopathological grade, TNM stage, tumour size and necrosis among the leading contributors, while age and performance status added little. Visual explanations showed the radiology branch attending to irregular tumour margins and heterogeneous enhancement and the pathology branch concentrating on sarcomatoid change and higher-grade regions. A combined view of feature importance indicated that radiomic and pathomic descriptors added predictive value beyond baseline clinical characteristics.
Error profiling highlighted where caution is warranted. False negatives were more frequent in tumours without necrosis, and false positives were more common in male patients, in stage II or III disease, in higher grades and in the presence of necrosis. Case-level reviews suggested areas for refinement, such as better attention to renal vein tumour thrombus on CT and more balanced weighting between central necrosis and enhancing peripheral tumour. The authors also noted practical boundaries. The analysis was retrospective, internal and external validation cohorts were drawn from Chinese populations, and imaging inputs were restricted to CT rather than MRI. Despite differing sample sizes across centres, external validation maintained clinically relevant discrimination.
An integrated score built from routine clinical data, contrast-enhanced CT and histopathological slides provided more reliable recurrence risk stratification for surgically treated stage I–III ccRCC than established clinicopathological tools or single-modality models. Strong performance in independent cohorts, a balanced sensitivity–specificity profile and correction of key misclassifications suggest value for guiding surveillance and adjuvant therapy decisions. Transparent behaviour, clear error patterns and stated limitations point to concrete avenues for broader validation and technical refinement, supporting cautious but pragmatic consideration of multimodal AI in everyday kidney cancer care.
Source: npj digital medicine
Image Credit: iStock
