Prognostic performance of machine learning in predicting haematological cancer outcomes: systematic review and meta-analysis

Diagnosis and treatment of haematologic malignancies present significant challenges, underscoring the need for highly individualized therapeutic approaches. Incorporating machine learning (ML) algorithms into predictions of haematological cancer outcomes has been increasingly investigated in recent years. However, it has not been investigated whether ML-based approach is superior to standard regression methods. Therefore, this review aims to assess their performance as compared to standard regression-based prediction methods. Studies from Web of Science, Medline, SCOPUS, CINHAL were reviewed, and associated risk of bias (ROB) assessed using PROBAST (Prediction Model Risk of Bias Assessment Tool). Standard and three-level random-effects meta-analysis were performed to obtain the ML pooled discriminative ability. Of 4204 retrieved studies, 48 were included in the systematic review. Pooled area under the curve (AUC) values of 24 top-performing ML and all 71 ML models from 19 studies were 0.80 (95% CI: 0.76, 0.84) and 0.779 (95% CI: 0.731, 0.827) in the testing set, respectively. The discrimination ability was similar between top-performing ML algorithms (AUC = 0.78, 95% CI: 0.70-0.86) and standard regression (AUC = 0.72, 95% CI: 0.66-0.78) in testing set. The three-level meta-analysis that incorporated all ML algorithms revealed similar results. However, externally validated top-performing ML algorithms outperformed standard models with a pooled AUC of 0.87 (95% CI: 0.76-0.98) compared with 0.72 (95% CI: 0.66-0.79). Although ML models' performance was acceptable, studies generally exhibited high ROB, emphasizing the need for future ML models to adhere to PROBAST guidelines.