Bladder cancer is the second most common urinary tract malignancy, yet early detection remains challenging. It ranks as the fourth most common cancer in men in the United States and carries a high recurrence rate after treatment. Despite its prevalence, there is no widely adopted non-invasive early screening tool comparable to PSA testing for prostate cancer or colonoscopy for colorectal cancer.
The current gold standard for diagnosis combines cystoscopy and urinary cytology. Cystoscopy is invasive, expensive, and carries procedure-related risks. Urinary cytology, while non-invasive, has poor sensitivity of only about 38% for low-grade tumors, leaving a large proportion of early cancers undetected until a more advanced stage.
Routine clinical laboratory tests are collected as standard of care for virtually every patient who presents for urological evaluation. These tests are inexpensive, widely available, and already part of existing clinical workflows. If machine learning models could extract diagnostic signals from these existing datasets, they could potentially serve as low-cost screening tools without requiring any additional testing.
This Taiwan-based study from Mackay Memorial Hospital investigated whether machine learning models trained on routine clinical laboratory values could predict bladder cancer and distinguish it from other urological conditions, including kidney cancer, prostate cancer, uterine cancer, and cystitis. The goal was to identify a practical, data-driven approach to early bladder cancer detection.
1,336 patients across five diagnostic groups provided the study dataset. The dataset was drawn from Mackay Memorial Hospital medical records and included patients with bladder cancer (591 patients), cystitis (144), kidney cancer (200), prostate cancer (201), and uterine cancer (200). The inclusion of four non-bladder-cancer groups allowed the model to be tested for specificity against both benign urological conditions and other urological malignancies.
A total of 56 clinical laboratory tests were initially collected for each patient, covering hematological, biochemical, and urinalysis parameters. All tests are part of routine clinical evaluation and do not require any specialized or experimental measurements. This makes the approach immediately applicable without requiring changes to standard diagnostic workflows.
Tests with more than 50% missing values were excluded from analysis, reducing the final feature set to 31 laboratory parameters. Missing values in the retained features were imputed using mean or median substitution depending on the distribution of each variable. This preprocessing step was critical because real-world clinical laboratory data rarely has complete entries for every test across all patients.
Class imbalance between the bladder cancer group and the smaller comparison groups was addressed using both oversampling and undersampling techniques. Oversampling generated synthetic minority-class samples to match the bladder cancer population size, while undersampling reduced the majority class to balance the dataset. Both approaches were evaluated and their performance compared.
A two-step feature selection process identified the most predictive laboratory markers. The first step used InfoGain with Ranker in the WEKA software platform to rank all 31 features by their information gain relative to the outcome variable. This step quickly eliminates non-informative features from consideration and provides an initial ranked list.
The second step applied forward selection, beginning with the highest-ranked feature and iteratively adding the next most informative feature, retaining it only if it improved model performance. This sequential approach identifies a compact, non-redundant feature subset that maximizes predictive power while avoiding overfitting from including too many correlated variables.
Five machine learning classifiers were evaluated: decision tree, random forest, support vector machine, XGBoost, and lightGBM. All models were trained and evaluated using 10-fold cross-validation, in which the dataset is divided into ten equal folds and each fold serves once as the test set while the remaining nine folds form the training set. This approach provides a reliable estimate of generalization performance.
Performance was measured using accuracy, sensitivity, specificity, and area under the ROC curve (AUC). For the primary clinical task of distinguishing bladder cancer from cystitis, the sensitivity metric was prioritized because false negatives (missed cancers) carry far greater clinical consequences than false positives in a screening context.
Eight laboratory features were selected as the final predictive panel. The two-step selection process converged on calcium, alkaline phosphatase (ALP), albumin, urine ketone, urine occult blood, creatinine, alanine aminotransferase (ALT), and presence of diabetes. These eight features represented the optimal combination for distinguishing bladder cancer from cystitis.
Urine occult blood is the most directly intuitive biomarker in this panel, as hematuria is a well-recognized symptom of bladder cancer. Its inclusion validates the biological plausibility of the feature selection process. The presence of occult blood in urine can indicate urothelial disruption from tumor growth even before macroscopic hematuria develops.
Calcium and ALP are both associated with bone metabolism and can be elevated in cancers with bone involvement, but they also reflect systemic metabolic changes associated with malignancy. Albumin is a marker of nutritional status and systemic inflammation, and low albumin is associated with poor outcomes across multiple cancer types. Its selection as a predictive feature likely reflects the metabolic burden of bladder cancer.
Creatinine reflects kidney function, which can be affected by urinary obstruction from bladder tumors. ALT reflects hepatic function, and its inclusion may capture systemic metabolic changes associated with malignancy. The inclusion of diabetes as a binary feature is consistent with epidemiological evidence linking diabetes to increased bladder cancer risk, possibly mediated through insulin signaling and bladder epithelial cell proliferation.
LightGBM achieved the best performance for the primary classification task. Among all five classifiers evaluated on the bladder cancer versus cystitis comparison, lightGBM produced the highest overall performance with accuracy of 87.6%, sensitivity of 89.5%, specificity of 85.5%, and an AUC of 0.932. These results suggest the model can reliably identify bladder cancer patients presenting with symptoms that might otherwise be attributed to benign cystitis.
The clinical importance of the bladder cancer versus cystitis comparison cannot be overstated. Both conditions present with hematuria and lower urinary tract symptoms, making clinical differentiation difficult without invasive testing. A non-invasive model achieving 89.5% sensitivity and 85.5% specificity could serve as a valuable triage tool to prioritize patients for cystoscopy.
Random forest and XGBoost also performed well, with AUC values above 0.90 in multiple comparisons. Decision tree and SVM showed lower performance, consistent with the known tendency of ensemble methods to outperform simpler single-model approaches on small, noisy clinical datasets.
The lightGBM model's superior performance likely reflects its gradient boosting architecture, which builds an ensemble of weak learners sequentially, each correcting the errors of the previous one. Its ability to handle missing data natively and its computational efficiency on small-to-medium datasets make it particularly well-suited to clinical laboratory data applications.
LightGBM consistently distinguished bladder cancer from other cancer types with AUC values of 0.88 to 0.92. When evaluated against kidney cancer, prostate cancer, and uterine cancer separately, the lightGBM model maintained strong discrimination performance across all pairwise comparisons. This cross-cancer specificity is clinically important because misdiagnosis between urological malignancies can lead to incorrect treatment selection.
The bladder cancer versus kidney cancer comparison is particularly clinically relevant because both conditions can present with hematuria. The model's ability to distinguish between them using routine lab values alone, without imaging, could assist in directing the diagnostic workup toward the correct anatomical site.
The bladder cancer versus prostate cancer distinction is similarly important in older male patients, where both conditions are common and can co-exist. Distinguishing their relative contributions to urinary symptoms and laboratory abnormalities using an automated tool could streamline the diagnostic process and reduce time to appropriate specialist referral.
The consistency of AUC values across all cancer-versus-cancer comparisons (0.88 to 0.92) suggests that the selected eight-feature panel captures features of bladder cancer biology that are generalizable across different comparison groups, rather than overfitting to characteristics specific to a single comparison group.
Routine laboratory data contains meaningful predictive signals for bladder cancer diagnosis. This study demonstrates that a compact panel of eight commonly available laboratory tests, analyzed with an optimized lightGBM model, can achieve clinically meaningful discrimination between bladder cancer and both benign urological conditions and other malignancies. No specialized tests, imaging, or invasive procedures are required.
The most immediate clinical application is as a pre-cystoscopy screening tool. Patients presenting with hematuria or lower urinary tract symptoms could be evaluated using the eight-feature model before being scheduled for cystoscopy. High-risk model outputs would prioritize patients for earlier cystoscopy, while low-risk outputs could support a watchful waiting strategy in resource-limited settings.
Study limitations include the single-center design from a Taiwanese hospital, which may limit generalizability to populations with different demographic and disease characteristics. The relatively small size of the cystitis group (144 patients) compared to the bladder cancer group (591 patients) also limits the precision of the bladder cancer versus cystitis comparison despite the oversampling correction.
Future work should validate the model in external multi-center cohorts, explore whether additional biomarkers beyond the eight selected features can further improve performance, and investigate whether the model adds value when combined with urinary cytology or imaging findings. Prospective clinical trials evaluating the model's impact on time to diagnosis and clinical outcomes would provide the strongest evidence for routine adoption.