Explainable AI Reveals What Shapes Recovery After Prostate Cancer Surgery

Artificial intelligence may be able to do more than predict what happens after prostate cancer surgery; it can also help doctors understand which factors are shaping those predictions. A new study, 'Explainable artificial intelligence reveals key surgical parameters in robot-assisted and open radical prostatectomy', published in nature in Scientific Reports, explored whether routinely available information collected before surgery could predict a range of surgical and postoperative outcomes while making the reasoning behind the AI models more transparent.

Researchers studied 326 men who underwent radical prostatectomy at University Hospital Bonn in Germany between 2015 and 2020. Of these patients, 224 underwent robot-assisted radical prostatectomy (RARP), while 102 received open radical prostatectomy (ORP). Rather than focusing on a single outcome, the team examined 12 targets covering blood-related changes, hospital stay, catheter duration, pathological findings and surgical decisions, creating a broader picture of where machine learning could realistically support preoperative planning.

Four AI Models Put Preoperative Patient Data to the Test

The researchers tested Random Forest, Gradient Boosting, Support Vector Machine and Neural Network models, using information that doctors could generally know before surgery. The available features included age, body mass index, PSA, preoperative hemoglobin and hematocrit, tumor stage, preoperative ISUP grade, Gleason score, prostate volume, D'Amico risk classification, surgical approach and planned nerve preservation. Features with substantial missing information or problematic overlap with other variables were removed before model training.

Because the study involved only 326 patients, the researchers used nested five-fold cross-validation, allowing them to tune the models while separately testing how well they performed on data not used during training. The models were also compared with simple baseline predictions and stronger statistical reference models, including Ridge regression, giving the researchers a clearer sense of whether machine learning was capturing useful information rather than simply reproducing obvious averages.

A Neural Network reached an R² of about 0.57, meaning the model explained a meaningful portion of the variation in postoperative hemoglobin levels. Hospital length of stay reached an R² of up to 0.46, while postoperative ISUP grade reached up to 0.37. Catheter dwell time and intraoperative fluid administration proved considerably harder to predict, generally producing R² values below 0.20.

The clearest classification result involved predicting whether surgeons would perform an intraoperative frozen section. Random Forest achieved an AUC of 89.2%, followed by Neural Network at 87.8%, Gradient Boosting at 87.4% and Support Vector Machine at 84.4%. The close results across very different algorithms suggest that this decision carries a relatively strong and structured signal in the available preoperative information.

Surgical Approach Emerged as a Powerful Predictor

The team applied an explainable AI technique based on SHAP, or SHapley Additive exPlanations, to identify how strongly individual preoperative features contributed to model performance. The approach allowed the researchers to rank predictive factors while stressing that these rankings describe associations learned by the models, not proof that one factor directly causes an outcome.

A striking divide appeared between outcomes shaped largely by surgical procedures and those more closely connected with patient or tumor characteristics. Surgical approach was responsible for more than 78% of predictive importance for catheter dwell time. RARP was also an important predictor of shorter hospital stays, while age and BMI contributed comparatively little to the prediction of hospital stay in this dataset.

The pattern extended to blood loss and postoperative blood measurements. Surgical approach accounted for 49.2% of predictive importance for intraoperative blood loss, with RARP associated with lower predicted blood loss than open surgery. RARP was also the leading factor for predicting hemoglobin drop, accounting for 31.5% of importance, while postoperative hemoglobin and hematocrit predictions drew strongly from both the surgical approach and patients' respective preoperative blood measurements.

These findings do not establish that robotic surgery itself caused each better predicted outcome. RARP and ORP patients were not identical groups, and differences in patient selection and hospital protocols could influence the patterns captured by the models. The researchers repeatedly caution against interpreting SHAP importance as a causal treatment effect.

Hemoglobin and Tumor Features Reveal Different Sides of Surgical Risk

Hemoglobin emerged as a strong predictor of postoperative blood-related outcomes and ranked above non-modifiable factors such as age when predicting hemoglobin drop. This raises the possibility that explainable machine learning could eventually help identify patients who may benefit from closer attention to anemia and blood management before an operation, though the researchers say prospective studies are needed before such predictions can guide treatment.

Postoperative ISUP grade was driven primarily by preoperative ISUP grade and biopsy Gleason score, which is clinically expected and gives the researchers a useful check that the models were learning established relationships. Preoperative ISUP grade contributed 18.5% of predictive importance, while planned nerve sparing contributed 17.3%. The authors caution that nerve sparing should not be interpreted as changing tumor biology; the surgical plan itself contains information from the surgeon's earlier assessment of tumor stage, imaging, biopsy findings and other risk indicators.

MRI-measured prostate volume was the strongest predictor of final prostate weight, contributing 27.5% of importance. The number of harvested lymph nodes reflected both surgical method and oncological risk, with RARP accounting for 39.9% of predictive importance, followed by factors including D'Amico risk score and initial PSA. Frozen-section predictions were influenced strongly by nerve-sparing plans and surgical approach, suggesting that AI was partly capturing the complex clinical judgement already embedded in the surgeon's planned strategy.

Promising Results Still Need Testing Beyond a Single Hospital

The research also shows where AI struggled, as some surgical complications and longer-term functional outcomes could not be predicted reliably because events were uncommon or categories contained too few patients. Higher-grade categories in the five-point Clavien-Dindo complication score were particularly sparse, preventing the models from performing substantially better than baseline predictions.

A model using preoperative hemoglobin alone explained about 26% of postoperative hemoglobin variation, compared with up to 57.5% for the machine-learning models, showing a meaningful improvement from using multiple variables. Multivariable Ridge regression performed similarly to machine learning for several outcomes, indicating that part of the useful predictive signal was accessible through simpler statistical approaches.

The study remains a single-centre, retrospective proof of concept without external validation. Differences between the RARP and ORP groups, possible unmeasured confounding, the relatively small sample, the handling of nerve-sparing plans and the approximate nature of the SHAP framework all limit how widely the findings can be applied. Calibration and decision-curve analyses were also not performed for the classification models.