Early Detection of Bladder Cancer Using Advanced Feature Engineering and Swarm Intelligence Optimization on EHRs.

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Bladder cancer (BC) symptoms often mimic benign conditions, complicating early diagnosis.
• There is a lack of routine screening for bladder cancer in the general population.
• Electronic Health Records (EHRs) offer a potential data source for developing screening tools.

Purpose of the Study:

• To develop and validate a machine learning (ML)-based screening pipeline for early bladder cancer (BC) detection.
• To utilize EHR data from primary care settings for BC screening.
• To identify novel predictors for BC risk through ML analysis.

Main Methods:

• A multi-centered case-control cohort (n=64,884) was used for model training and testing (1995-2018).
• Prospective validation was performed on an independent cohort (n=4,569) (2019-2020).
• The Parsimony driven REweighting for Calibrated Input-based Screening for Early detection-Adjustable Grey Zone (PRECISE-AGZ) pipeline identified influential features and developed a calibrated logistic regression model.

Main Results:

• The model identified 38 features, achieving an AUC of 0.789 on the testing set.
• Novel predictors included neurological disorders (e.g., Parkinson's disease) and medications (e.g., Tamoxifen).
• The low-risk category (predicted probability ≤0.55) achieved 85.2% sensitivity, 0.799 F1-score, and 34.5% screening population coverage.

Conclusions:

• The PRECISE-AGZ pipeline effectively extracts clinical signals from EHRs for early BC detection.
• This approach shows significant potential for implementing population-based bladder cancer screening.
• The study highlights the utility of ML in identifying subtle indicators for early disease detection.