gms | German Medical Science

Introduction: Prostate cancer is the second most commonly diagnosed cancer in men worldwide and one of the most leading causes of death in Western countries. After definitive primary treatment the cancer recurs in up to 30% of all patients. Ga-68-PSMA PET/CT has become an important method for additional diagnostics in recurring patients. While it performs great, it is an expensive, invasive and time consuming examination. Therefore, we employed modern state-of-the-art multivariate machine learning (ML) methods on electronic health records (EHR) of prostate cancer patients to improve the prediction of imaging confirmed prostate cancer recurrence (IPCR).

Methods: Our dataset consists of 240 patients, (208 recurring, 32 non-recurring). The ten most prevalent features (<20% missing values) from the EHR were used. To deal with imbalance and missing values, different preprocessing techniques such as upsampling and imputation were used. For the predictive analysis, we selected four different well-established ML models, namely Logistic Regression (LR), Random Forest (RF), Gradient Boosting (GB), and Support Vector Machine (SVM). The predictive performance was evaluated using sensitivity and specificity, the area under the receiver operating characteristic curve (AUC), the area under the precision-recall curve (AUCPR), and the Matthews Correlation Coefficient (MCC).

Results: The LR, RF, and GB models performed equivalently well, while the SVM performed more poorly. The highest mean in AUC (0.78 ± 0.13, mean ± standard deviation) was achieved by the RF model employing upsampling. All models are very sensitive to patients with IPCR (sensitivity ≤ 0.92 ± 0.06 for all models). However, all models perform fairly poorly on identifying individuals without prostate cancer recurrence, resulting in low MCC and specificity scores.

Discussion: For the prediction of PCR with ML methods, several studies focus on the prediction of biochemical recurrence (BCR) [1]. Since the detection rates of PCR based on BCR alone can be low for small PSA values [2], we opted to predict the IPCR determined by Ga-68-PSMA PET/CT imaging. This makes our task more difficult, mainly because Ga-68-PSMA PET/CT imaging is solely performed when recurrence is highly likely. The AUC score of our RF model is slightly higher than the convolutional neural network of Sargos et al. [3], that achieved an AUC score of 0.77 on BCR prediction after three years of primary treatment. Moreover, our model is only slightly worse than Lee et al. whose GB classifier achieved an AUC score of 0.8031 for BCR after five years of primary treatment [4].

Conclusion: While the performance of our pilot study clearly indicates the relevance of multimodal ML for PCR prediction prior to imaging, further analysis on larger cohorts will be needed to validate the presented findings. Additional analyses of specificity and MCC scores on a larger and more balanced cohort are required. In summary, our results demonstrate that adequate multimodal ML workflows can aid the prediction of IPCR and indicate patients with early-stage recurrence. This will reduce the amount of harmful PET/CT scans performed on patients and save costs and resources in the clinics.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.