OBJECTIVES: Machine-learning (ML) consist of developing algorithms that can generate output predictions based on learning from input data. This review aims to synthesize available studies comparing ML to traditional methods in the prediction of disease outcomes and diagnosis.

METHODS: A systematic literature review was conducted, looking at interventional studies comparing the performance of ML to traditional statistical methods and/or the performance of different ML methods on the ability to predict diagnosis and/or disease outcomes. Studies were included if they assessed any comparison of the application of different machine-learning models (with traditional methods or other type of machine-learning models) in diagnosis or outcomes prediction.

RESULTS: In total, 19 studies were identified, 12 comparing machine-learning models with traditional methods while 7 studies compared different machine learning models. 11 studies focused on applying ML methods in the prediction of health outcomes and 8 on machine-learning models as diagnostic methods. Therapeutic area with most studied applications was neurology (n=4) followed by oncology and ophthalmology (n=3 each). Applications of ML methods were also assessed in immunology, hepathology, pulmonology, dermatology, critical care and infectious diseases. Random Forest models and Support Vector models were the most frequent type of ML models used, with 8 models each. All the studies comparing machine-learning models with traditional methods of diagnosis or outcomes prediction demonstrated that ML models achieved superior results in sensitivity (18.82% to 29.38%), specificity (17% to 28.08%) and accuracy (10% to 12%) as well as other specific outcomes.

CONCLUSIONS: The studies identified in this review demonstrated that ML models are associated with increased value in diagnosis and outcomes prediction. Further model development and training with larger datasets may improve the predictive power of machine-learning methods.