To create an algorithm assessing the risk of ovarian cancer in primary care using generalized additive2 model (GA2M) and traditional methods.

Objectives: We aimed to evaluate models designed to support the exploration and early detection of potential Ovarian Cancer (OC) using either Machine Learning (ML) techniques or traditional methodologies, using primary care data. This evaluation aimed to facilitate appropriate and timely referrals to specialists.

Methods: The Health Search database, containing healthcare records of 1 million adults, was used to predict OC cases among patients aged 18+ without prior OC diagnosis from 1 January 2002, to June 2021. GA²M, logistic regression, and Gradient Boosting Machines (GBM) were trained using 20 determinants, with prediction performance assessed via Area Under Curve (AUC) and Average Precision (AP). A traditional cohort design with nested case-control analysis developed a prediction algorithm using the same determinants, evaluating explained variation, AUC, and slope calibration.

Results: Comparing the predictive performances of the three models, the AUC and AP for GA²M showed the highest values, which were equal to 69.1 and 0.7%, respectively. The GA²M-based algorithm outperformed the algorithm obtained with the traditional approach, which showed an overestimation of risk, as confirmed by a calibration slope of 1.75 along with an AUC of 55%.

Conclusions: The GA²M-based algorithm outperformed the light GBM, logistic regression, and traditional models in predicting OC. This suggests that ML techniques are preferable for algorithms involving complex predictor interplays in assessing rare event risks, such as OC. The GA²M-based algorithm is reliable for OC prediction in primary care compared to traditional methods, indicating its potential for use in decision support systems for general practitioners.