Session Item

To develop a simulated annealing optimization algorithm for building Bayesian Network (BN) topologies for late toxicity prediction in the EMBRACE I trial on locally advanced cervical cancer after radiochemotherapy.

EMBRACE morbidity outcomes were used to develop an algorithm for building two separate BN topologies for grade ≥2 cystitis and proctitis, as assessed by CTCAEv3. Baseline and at least one follow-up assessment were required. 40 clinical and treatment features were included as potential network nodes in the optimization. All features were initially included to ensure a direct comparison between the different model-building techniques. Restrictions on node connectivity were specified by experts to ensure that connections followed a logical sequential order (i.e. dosimetric features from treatment cannot be parent nodes of diagnostic features). Topology optimization followed a simulated annealing framework, with the cost function containing terms prioritizing model performance, mutual information between nodes, and simplicity. Results of the simulated annealing optimizations were compared to out-of-box optimization algorithms from the PyAgrum package (Tree-Augmented Naïve Bayes (TAN), Greedy Hill Climbing (GHC), Chow-Liu Optimization (CLO)).

The proctitis model was built using 1108 patients and the cystitis model was built with 1073, excluding patients with bladder wall infiltration. Simulated annealing optimization for both toxicities terminated with convergence following 1000 unchanged iterations. Results for model performance and topology characteristics are outlined in Table 1 and networks are shown in Figure 1.

Simulated annealing resulted in models with comparable predictive accuracy to out-of-box optimization techniques (Cystitis: SA=57.3±6.6%, TAN=53.7±16.2%, GHC=55.4±6.4%, CLO=56.4±6.5%. Proctitis: SA=55.2±7.0%, TAN=42.4±14.5%, GHC=53.8±5.1%, CLO=53.3±5.8%). The benefit of the simulated annealing technique is the highly tunable nature of the optimization, allowing for a customizable cost function and expert-specified restrictions placed on connections reducing the of the solution space. As a result, the simulated annealing framework notably resulted in a less complex network than the TAN technique and did not suffer from illogical connections that other models encountered.

Simulated annealing-based optimization resulted in BN topologies with comparable predictive accuracies to other out-of-box optimization techniques. The simulated annealing approach constructed more interpretable and logical topologies while allowing for customization within the cost function and initialization settings to allow for the inclusion of clinical knowledge in the optimization process. Although model accuracy is currently limited, this technique provides a preliminary approach for late toxicity prediction, which remains a complex task with no standard methodology for identifying at-risk patients.