Figure 1.
The experimental setup for structure learning.
The pseudo-code of the experimental setup for learning and assessing DAGs via different algorithms.
Figure 2.
Schematic explanation of a causal intervention on a BN.
Figure 3.
DAG structures learned by different methods.
(a) manual construction, (b) Tree Augmented Naive Bayes (TAN) algorithm, (c) CaMML algorithm with structural pair-wise priors, (d) CaMML algorithm with temporal tiers.
Figure 4.
Expert elicited structural pairwise relations based on the selected 13 LUCADA variables.
The variable codes are as given in Table 1. The notations can be read as: “A≺B”: A happens before B; “A−B”: A and B are related; and “A→B”: A influences B.
Figure 5.
The treatment plan probabilities as calculated from LUCADA.
P(Treatment) is represented in striped blue columns, and the conditional 1-year survival probabilities given specific treatment plans, P(Surv = Alive│Treatment), is represented in green columns The horizontal axis, which contains the treatment plan options, is ordered in descending order of P(Treatment) from left to right.
Figure 6.
The modified BN Structure for causal interventions.
All edges towards the intervened “Suggested Cancer Treatment Plan” are removed.
Figure 7.
The confusion matrix that displays the recorded versus the recommended treatment plans.
The recommended treatment plans are the ones that maximise 1-year survival for a patient, acquired via causal intervention on the BN.
Figure 8.
The exact and partial concordances between the recommended and the recorded treatment plans.
The concordances are stratified with respect to treatment plan types. The recommended treatment plans are the ones that maximise 1-year survival for a patient, acquired via causal intervention on the BN.
Figure 9.
The non-surgical confusion matrix for patients who have been treated with non-surgical treatment plans.
The recommended treatment plans are the ones that maximise 1-year survival for a patient, acquired via causal intervention on the BN.
Figure 10.
The exact and partial concordances between the recommended and the recorded treatment plans.
The concordances are stratified with respect to the TNM stages.