Application of Machine learning in predicting cancer complications using longitudinal Data: A systematic review and...

Cancer prognosis of complications like metastasis, recurrence, and side effects of treatments is important to enhance patient prognosis. There is great potential in the use of ML on lifetime data for improving prediction accuracy in oncology; however, there is no systematic review of the subject. This SRMA is intended to assess the accuracy of ML models based on longitudinal studies for the estimation of cancer-related complications. The articles were identified from PubMed, Google Scholar, and IEEE Xplore databases for the years 2020 to 2024. Seven of the studies reviewed in the paper analyzed ML models that employed longitudinal data for cancer complication prognosis. The risk of bias of included studies was assessed using the Cochrane Risk of Bias tool, and for diagnostic accuracy, the QUADES 2 tool was used. Information on ML techniques, prediction accuracy, and results was obtained. The pooled area under the curve (AUC) for immune-related adverse events prediction was 0.78 (95% CI: 0.73-0.83). For cancer recurrence and mortality prediction, pooled AUCs ranged from 0.70 to 0.75. Machine learning models integrating clinical, genomic, and imaging data demonstrated superior predictive accuracy across various cancer types. Models predicting quality of life deterioration during treatment showed an AUC of 0.82. ML models applying longitudinal data effectively predict cancer complications with improved accuracy when integrating multimodal data. These models offer promising tools for clinical decision-making in oncology.