Active surveillance (AS) is the recommended initial management for most men diagnosed with very-low or low-risk prostate cancer. Rather than immediately treating a cancer that may never cause harm, AS involves close monitoring -- with regular PSA tests, physician exams, and repeat biopsies -- to detect any signs of cancer progression before it becomes dangerous.
Long-term data have confirmed that AS can achieve excellent outcomes for favorable-risk disease. However, the monitoring protocol itself carries burdens: repeat prostate biopsies carry risks of bleeding, urinary symptoms, erectile dysfunction, and infection (including severe infections requiring hospitalization). The process is also time-consuming and costly for patients and healthcare systems.
Beyond physical risks, some men on AS experience cancer-related anxiety -- the psychological burden of living with a diagnosed cancer while deferring treatment. This anxiety can reduce quality of life and make patients more likely to abandon AS in favor of surgery or radiation, even when continued monitoring is medically appropriate.
A robust tool that could identify, at diagnosis, which patients are most likely to experience cancer progression -- and which patients are very unlikely to progress -- could enable personalized surveillance protocols: fewer biopsies for low-risk patients, more vigilant monitoring for high-risk patients. Despite this need, all existing models to predict AS progression had been developed using traditional statistical approaches, which have important limitations.
The study retrospectively analyzed 790 patients diagnosed with very-low or low-risk prostate cancer (PSA less than 10 ng/mL, clinical stage up to T2a, and ISUP grade group 1) who enrolled in an AS program at Massachusetts General Hospital between 1997 and 2016. The target outcome was grade progression -- detection of ISUP grade group 2 or higher cancer on any subsequent biopsy, signaling that the cancer had become more aggressive.
The dataset was split 80/20 into a training set (632 patients) and a test set (158 patients) using stratified sampling to preserve the proportion of progressors in each group. With a median follow-up of 6.29 years, 234 patients (30%) experienced grade progression -- creating a class-imbalanced dataset where the event of interest (progression) was the minority class.
Five predictive models were compared: (1) traditional logistic regression (T-LR) using automated backward variable selection, (2) machine learning logistic regression (ML-LR), (3) support vector machine (SVM), (4) random forest, and (5) a fully connected artificial neural network (ANN). Input features included age, PSA level, PSA density, prostate volume, biopsy core characteristics, family history, clinical stage, use of 5-alpha reductase inhibitors, and surveillance biopsy history.
The primary performance metric was the F1 score -- the harmonic mean of precision and sensitivity. This metric is preferred over simple accuracy in class-imbalanced datasets because it equally penalizes both missing true progressors (false negatives) and incorrectly flagging stable patients as progressors (false positives). A perfect F1 score is 1.0. The ML models were tuned specifically to optimize F1, while the T-LR model used standard accuracy-based variable selection.
All four ML models significantly outperformed the traditional logistic regression model (p less than 0.001 for all comparisons). The support vector machine achieved the highest F1 score of 0.586, followed by ML-LR at 0.522, artificial neural network at 0.392, and random forest at 0.376. The traditional logistic regression scored a markedly lower F1 of 0.182.
The traditional logistic regression model's low F1 score reflected a fundamental problem: it overwhelmingly predicted that patients would NOT progress (the majority class), achieving 97.3% specificity but only 10.6% sensitivity. In practice, this means it correctly identified stable patients but missed nearly 9 out of 10 true progressors -- clinically unacceptable for a surveillance tool.
The best ML model (SVM) achieved a sensitivity of 72.3% and specificity of 68.5%. Its negative predictive value was 85.4%, meaning that when the SVM predicted no progression, the patient was truly stable about 85% of the time. This is the clinically meaningful performance characteristic -- it gives physicians confidence that those classified as low-risk can genuinely reduce surveillance intensity.
The T-LR final model retained only 4 variables after backward elimination: age, PSA density, maximum percentage of core involvement, and time between initial diagnosis and most recent biopsy. The ML models used all available features without selection, allowing them to capture complex interactions between variables that the step-wise elimination approach would miss or discard.
The dramatic gap in sensitivity between traditional logistic regression and ML models can be traced to how each method handles class imbalance. When only 30% of patients progress, a model optimized for overall accuracy can appear highly accurate by simply predicting no progression for everyone. Traditional logistic regression, using accuracy-based variable selection, fell into this trap. ML models were explicitly tuned to optimize the F1 score, forcing them to balance precision and sensitivity rather than maximizing easy gains from the majority class.
Traditional logistic regression also assumes linear relationships between predictors and the outcome. ML algorithms like SVMs and neural networks can learn non-linear, interaction-based patterns -- for example, the combined effect of PSA density and core involvement percentage may be more predictive than either factor alone, and this synergy would be lost in a simple linear model.
The authors noted that comparing models using the c-statistic (area under the ROC curve) can be misleading in imbalanced datasets. The T-LR model achieved a c-statistic of 0.686 -- actually higher than the random forest (0.603) and neural network (0.545) -- despite its clinically inadequate sensitivity of 10.6%. The c-statistic in such settings gives an overly optimistic impression of performance because it equally weights sensitivity and specificity even when sensitivity matters most clinically.
Comparison with previously published AS progression models revealed that c-statistics from those studies (typically 0.70-0.76) are difficult to interpret at face value. The present study's SVM achieved a c-statistic of 0.701 -- similar to those prior models -- but this masks a substantially better sensitivity profile, because the SVM was optimized for a different goal. This highlights the importance of choosing evaluation metrics that reflect actual clinical priorities.
A clinically effective AS progression model would enable risk-stratified surveillance: patients identified as low-risk for progression could follow a less intensive schedule (fewer biopsies, longer intervals between PSA checks), while high-risk patients would undergo more vigilant monitoring. This approach would reduce biopsy-associated harms for the majority of patients who will never experience progression.
The SVM model's negative predictive value of 85.4% is particularly relevant for de-escalation decisions. A patient classified as low-risk by the model has approximately an 85% chance of truly being stable over the surveillance period -- enough confidence to consider extending the biopsy interval without exposing the patient to unacceptable risk of missing a meaningful progression event.
Beyond reducing physical risks, a more reliable prediction tool could alleviate cancer-related anxiety. Evidence shows that long-term anxiety is common among AS patients. If men with very low predicted progression risk could receive a data-driven reassurance that their cancer is unlikely to advance, some may experience reduced psychological burden and greater willingness to continue surveillance rather than opting for unnecessary treatment.
Future model improvements should incorporate MRI features. Prostate multiparametric MRI (mpMRI) is increasingly used in AS and provides information beyond what biopsy and PSA alone can offer. Combining MRI data with clinical variables through convolutional neural networks -- as suggested by the authors -- could substantially improve both sensitivity and specificity beyond what was achieved in this study.
This is the first study to apply machine learning to the prediction of prostate cancer progression on active surveillance. The results demonstrate that ML methods, particularly the support vector machine, substantially outperform traditional logistic regression when the clinical goal is identifying true progressors in an imbalanced patient population.
Important limitations include the retrospective, single-institution design and the absence of MRI data. External validation in independent cohorts is critical before clinical deployment, as model performance may vary across populations with different clinical characteristics or institutional follow-up practices. The authors acknowledge that traditional statistical approaches could potentially be competitive in some datasets.
Nonetheless, this study provides strong impetus for developing larger, more feature-rich ML models for AS management. The ideal next step would be a prospective multicenter study incorporating MRI features, molecular biomarkers, and longer-term outcomes to develop and validate a clinically deployable prediction tool.