Breast lesion risk prediction

Breast cancer is one of the leading causes of mortality among women and has high potential for cure when detected early. In private health insurance, Primary Care can support care coordination, risk factor identification and better targeting of prevention, screening and diagnostic pathways.

The payer observed frequent and costly use of mammography and breast ultrasound, including among young patients without identified risk factors and without later positive diagnoses. At the same time, it was necessary not to miss beneficiaries at increased risk, especially outside traditional screening criteria.

Develop a prediction and risk stratification algorithm for breast problems, supporting individualized Primary Care pathways for prevention, early diagnosis and more targeted referral.

The project used real private datasets from a health insurance organization, including: - Electronic health records - Self-administered questionnaires - ERP and medical claims data - Exam results - Gym access logs from partner gyms

The datasets were integrated to build a final risk factor matrix. Twenty-one Boolean features were identified, representing the presence or absence of factors associated with breast problems. The target grouped benign and malignant breast alterations because isolated malignancy had low occurrence in the sample.

The factors considered included: - Family and personal history of breast cancer - Ovarian neoplasia - Biopsy history - Breast-related diagnoses - Breast density - Obesity - Smoking - Alcohol use - Sedentary behavior - Inadequate diet - Hormonal and gynecological-obstetric factors - Other relevant clinical conditions

Different machine learning algorithms were compared: - Logistic Regression - Random Forest - Support Vector Machines - Sequential Neural Network The comparison used cross-validation, hyperparameter optimization and evaluation by area under the ROC curve.

The best model was a Logistic Regression with the following hyperparameters: - class_weight: balanced - max_iter: 1000 - penalty: l2 This choice favored a balance between performance, interpretability and applicability in a healthcare decision support context.

In the demonstration app, the confusion matrix shows: - TN: 1,264 - FP: 737 - FN: 9 - TP: 49 These results reinforce that the operating point must balance sensitivity and precision according to care capacity. In prevention and screening, threshold selection should consider the cost of false negatives, the cost of false positives and care availability.

The technical report notes that published breast cancer risk prediction models using epidemiological, demographic and clinical data reported a mean AUC around 0.73, with 95% CI from 0.66 to 0.80. The developed model reached a mean AUC of 0.8165, suggesting competitive performance in an applied setting using real payer datasets.

The model should be interpreted as a decision support tool, not as a replacement for clinical judgment. Its proposed use is to support prioritization of beneficiaries for Primary Care pathways, clinical investigation, preventive guidance, individualized screening and eventual diagnostic referral.

- The target grouped benign and malignant alterations, reducing clinical specificity for isolated malignant cancer. - The low occurrence of malignancy required a broader breast problem modeling approach. - Data came from a specific payer and require external validation. - The model should be periodically recalibrated. - The operating point should be defined according to care capacity and tolerance for false positives and false negatives. - The tool does not replace clinical assessment, care guidelines or shared decision-making.

- Demonstration web application - Logistic Regression model - Feature matrix with 21 Boolean variables - Comparison between machine learning algorithms - ROC curve - Precision-Recall curve - Confusion matrix - Hyperparameter evaluation - Technical project material - Technical presentation

- Health risk prediction needs to be connected to a clear care pathway. - In screening, threshold selection is as important as the global metric. - Real payer datasets require integration of multiple sources and well-defined business rules. - Grouping benign and malignant events may be necessary when isolated malignancy occurrence is low, but it changes the clinical interpretation of the model. - The greatest value lies in supporting care coordination, not automating diagnosis.