Machine learning—enhanced baseline risk prediction models for cancer therapy related cardiac dysfunction: a systematic review and critical appraisal of methodological quality

Figures

• 

  Figure 1. PRISMA flowchart of study inclusion

• 

  Figure 2. Characteristics of the included studies. (A) Type of primary malignancy diagnosed in the patients included in each study. (B) Type of cardiotoxic cancer therapy included. (C) Frequency of Machine Learning models reported across the 9 studies and 28 models. (D) Different types of data used as predictor variables in each study,

• 

  Figure 3. Summary of the reported results. (A) Forest plot of the area under the receivers operator characteristics curve reported for each study and the confidence intervals (CI) wherever reported. Tsiouris et al.’s18 model is not included in this graph because an AUC was not reported. He et al.21 and Chang et al.19 are represented as a dot without distribution bars because the CI were not reported. If the study included more than one ML model, the best-performing algorithm is represented in this forest plot. (B) Bargraph showing the frequency of appearance as the most influential variable across all models. Each study identified and reported the most impactful predictors influencing their model performance. Of the clinical features, hypertension and LVEF emerged as the most frequently significant variable.

• 

  Figure 4. PROBAST risk of bias assessment results of the nine studies. Domains: D1: Participants—Assesses bias due to participant selection, recruitment, or inclusion/exclusion criteria. D2: Predictors—Evaluates bias arising from the measurement or classification of predictors. D3: Outcomes—Assesses bias in outcome measurement, including objectivity and measurement consistency. D4: Analysis—Evaluates bias in statistical analyses, including handling of missing data, model appropriateness, and selective reporting. Overall—Summary risk of bias considering all domains. Judgment categories: Low (green ‘+’): Low risk of bias; methods and reporting are appropriate and unlikely to meaningfully distort results. Unclear (yellow ‘–’) : Risk of bias is uncertain due to incomplete information or methodological ambiguity. High (red ‘×’): High risk of bias; flaws in study design, conduct, or reporting are likely to distort results.

• 

  Figure 5. TRIPOD + AI percentage of reported items across studies. The percentages presented in this bar graph are a result of the number of reported items over the number of applicable items from the TRIPO + AI checklist to the particular study.