Prediction of metastatic relapse in intermediate/high risk localized prostate cancer patients from staging medical images and clinical variables (PROVIDENCE study)

Prostate cancer (PCa) is the second most common cancer in men worldwide and the second leading cause of cancer death in men. Most localized PCa represents an indolent disease; however, intermediate- and high-risk patients are more likely to die from prostate cancer and its prognosis is more difficult to predict. For these patients, prognosis remains complex with biochemical recurrence (BCR) marking a significant step towards disease progression and potentially metastases. Therefore, accurate risk stratification and swift diagnosis are imperative to enhance treatment decisions and decrease progression incidences.

This retrospective, multicenter study, PROVIDENCE, aims to develop an AI model predicting metastatic relapse in 300 intermediate or high-risk localized PCa patients. It utilizes diagnostic medical images including multiparametric MRI (T2-W, DWI, DCE sequences), staging bone scans, and clinical data. This is all consolidated on the QP-Insights^® platform. Automatic segmentation of prostate gland MRIs and lesion(s) detection followed by the extraction of quantitative imaging biomarkers (QIBs) is being performed, supplemented by counts quantification from staging bone scans. In conjunction with clinical variables, this data will inform the predictive model, forecasting patient metastatic relapse.

The poster presented in European Multidisciplinary Congress on Urological Cancers (EMUC) 2024 congress in Lisbon¹, showcased the first interim analysis, where clinical variables of 127 PCa patients were used to predict BCR. Both groups, the BCR (n=38) and no BCR (n=89), despite the imbalance, had a similar clinical distribution. Hence, 9 diagnostic clinical variables were included in the GLMnet model with 5-fold cross-validation for the BCR prediction. The model achieved an AUC of 0.79 (sensibility [se] =0.33, specificity [sp] = 0.39), showed in Figure 1. Surgery as the primary treatment for PCa and clinical stage were identified as the most significant variables influencing the model. In addition, by assessing the SHapley Additive exPlanations (SHAP) values, plotted in Figure 2, it can be concluded that undergoing surgery significantly lowered the risk of BCR, whereas elevated nadir prostate-specific antigen (PSA) levels, older age, and higher International Society of Urological Pathology (ISUP) grades increased the likelihood of BCR during follow-up.

Furthermore, a second interim analysis was performed with the results submitted to the European Association of Urology (EAU) congress²; these included the analysis of baseline MRIs from 164 patients. QP-Prostate^® was used to automatically segment the prostate gland, dividing it into 3 regions (central and transitional zone (CZ+TZ), peripheral zone (PZ) and seminal vesicles (SV)) and the lesion(s) within the prostate gland, as shown in Figure 3. Radiomic features were extracted from each segmentation mask and the volume-weighted average was computed for the whole prostate and the lesions. Furthermore, several machine learning (ML) models were assessed to predict the likelihood of metastatic recurrence by using the radiomics features of the whole prostate, of the volume-weighted average lesion and of the largest lesion. The corresponding metrics are shown in table 1. Data augmentation, by using Synthetic Minority Over-sampling Technique (SMOTE) method was done due to the imbalance among groups, 42 with metastatic recurrence and 122 without. The random forest (RF) model trained with radiomic features extracted from the volume-weighted average lesions demonstrated the highest performance (AUC 0.92).

A final analysis that will include a larger cohort and potentially more balanced data, is expected to yield more definitive insights into the predictive value of QIBs and clinical data for PCa prognosis. This would enable non-invasive real-time patient stratification,and thereby facilitate the customizing adjuvant therapy and the potential intensification of initial treatments for localized PCa patients diagnosed as intermediate or high risk.

References

1. EMUC Congress 2024: An interim analysis of the multi-center retrospective study PROVIDENCE: Artificial intelligence-based models for estimating the risk of biochemical recurrence with clinical data. EMUC24-0221. Published 2024. doi:10.1016/S2666-1683(24)01195-9.

2. EAU25 40th Annual EAU Congress (submitted abstract): An interim analysis of the multi center retrospective study PROVIDENCE: Artificial intelligence-based model for predicting metastatic recurrence in localized prostate cancer patients using quantitative imaging biomarkers. AM25-0767. Pending evaluation.