From Real-World Data to Causally Interpretable Models: A Bayesian Network to Predict Cardiovascular Diseases in Adolescents and Young Adults with Breast Cancer

Figures

• 

  Figure 1. Final Bayesian network model structure. The figure illustrates the directed acyclic graph behind the model. Briefly, the cohort of origin is set as the root context node, followed by three subsequent lines that illustrate the cancer prognostic factors which affect both the 5-year survival and treatment decisions (neoadjuvant if pre-surgery, or adjuvant if post-surgery). Type 2 diabetes (t2db), dyslipidemia, and hypertension (bottom right in the figure) can already be present prior to the cancer diagnosis or can be induced by the cancer treatments. Finally, at the bottom of the model, the nodes can be seen describing cardiotoxicity induced by treatments and ischemic heart diseases, pointing to the target variable (i.e., cardiovascular diseases (CVDs)).

• 

  Figure 2. Bayesian network classification performances. Each classification metric is represented by a bar and the error bars are the 95% confidence intervals. Generally speaking, the model had good classification performances (AUC = 88% and accuracy = 99%). The model showed an almost perfect ability to correctly predict survivors not developing a CVD as not at risk (TNR, NPV), while identifying those developing a CVD as at risk is more complex due to the outcome rarity (TPR). When class imbalance is taken into consideration (BA and MCC), the performances decrease but remain very high (>80%).

• 

  Figure 3. Lift chart obtained using Bayesian network model predictions. In this graph, survivors are sorted (x-axis), in decreasing order, according to their risk of developing a cardiovascular disease (CVD). While the cumulative distribution of the survivors increases, the corresponding True Positive Rate (i.e., the proportion of survivors in the validation set who actually developed a CVD within 5 years) is plotted on the y-axis. The sharp rise in the lift chart highlights the model’s good ability to identify survivors who should be prioritized for CVD follow-up.

• 

  Figure 4. Concept draft of the interface (a) and the output (b) of the app-based tool. The figure illustrates the interface of the desktop-based app based on Bayesian network model probabilities. When the user inputs the patient’s information into the data entry window (a), the baseline model probability estimates for the 5-year cardiovascular disease risk are updated. Subsequently, in a new window (b), the user is provided with the individual predicted patient probability, together with an illustration that helps the user interpret the predicted value according to the risk classes.