About The Project
Renal cell carcinoma (RCC) represents about 3 percent of all cancer-related cases at 2018, with the highest incidence occurring in Western countries. During the last decades, stage migration towards localized disease has occurred. Partial nephrectomy is the treatment of choice for localized cT1 renal masses. The main advantage of partial nephrectomy is the preservation of renal function compared to radical nephrectomy. One of the most devastating complications of partial nephrectomy is post-operative acute kidney injury (AKI), which increases the risk of long-term chronic kidney disease (CKD) with its consequences, including decreased overall survival. The prevalence of AKI following partial nephrectomy is reported to be up to 24.3 percent, and is dependent on surgical approach, patients baseline characteristics and the definition of AKI used in each study.
In this study, we apply explainable machine learning (ML) model to predict patients who would develop AKI following partial nephrectomy. We hypothesize that patterns from pre-surgery could be identified and learned by ML models. A self-explainable prediction system is based on ML was than built and deployed online to help doctors evaluate the risks in using the partial nephrectomy treatment on RCC patients, provide patients with individualized treatment.
Since 1995, we have been continuously updating our PN database to include clinical, surgical and oncological parameters. Currently, it includes 723 patients. For this particular study, we included all adult (> 18 years) patients who underwent PN for enhancing solid renal mass(es) and had all the data to be grouped as either post-operative AKI or non-AKI.
Afterward, we perform six steps (on the right) to obtain the model.
Data split according to the AKI value (due to imbalanced data)
We used the random forest algorithm (with majority vote).
Greedy feature selection (from empty set) with accuracy metric.
Performing SAT-based pruning.
Performing grid search on the tree's depth, minimal number of sample for leaf, the leaf count, and the number of trees in the forest.
Five-fold cross-validation on several metrics.
