ARCliDS: A Clinical Decision Support System for AI-assisted Decision-Making in Response-Adaptive Radiotherapy

Abstract

Background: Involvement of many variables, uncertainty in treatment response, and inter-patient heterogeneity challenge objective decision-making in dynamic treatment regime (DTR) in oncology. Advanced machine learning analytics in conjunction with information-rich dense multi-omics data have the ability to overcome such challenges. We have developed a comprehensive artificial intelligence (AI)-based optimal decision-making framework for assisting oncologists in DTR. In this work, we demonstrate the proposed framework to Knowledge Based Response-Adaptive Radiotherapy (KBR-ART) applications by developing an interactive software tool entitled Adaptive Radiotherapy Clinical Decision Support (ARCliDS). Methods: ARCliDS is composed of two main components: Artificial RT Environment (ARTE) and Optimal Decision Maker (ODM). ARTE is designed as a Markov decision process and modeled via supervised learning. Given a patient's pre- and during-treatment information, ARTE can estimate treatment outcomes for a selected daily dosage value (radiation fraction size). ODM is formulated using reinforcement learning and is trained on ARTE. ODM can recommend optimal daily dosage adjustments to maximize the tumor local control probability and minimize the side effects. Graph Neural Network (GNN) is applied to exploit the inter-feature relationships for improved modeling performance and a novel double GNN architecture is designed to avoid unphysical treatment response. Datasets of size 117 and 292 were available from two clinical trials on adaptive RT in non-small cell lung cancer (NSCLC) patients and adaptive stereotactic body RT (SBRT) in hepatocellular carcinoma (HCC) patients, respectively. For training and validation, dense data with 297 features were available for 67 NSCLC patients and 110 features for 71 HCC patients. To increase the sample size for ODM training, we applied Generative Adversarial Network to generate 10,000 synthetic patients. The ODM was trained on the synthetic patients and validated on the original dataset. Results: Double GNN architecture was able to correct the unphysical dose-response trend and improve ARCliDS recommendation. The average root mean squared difference (RMSD) between ARCliDS recommendation and reported clinical decisions using double GNNs were 0.61 ± 0.03 Gy/frac (mean ± sem) for adaptive RT in NSCLC patients and 2.96±0.42 Gy/frac for adaptive SBRT HCC compared to the single GNN's RMSDs of 0.97 ± 0.12 Gy/frac and 4.75 ± 0.16 Gy/frac, respectively. Overall, For NSCLC and HCC, ARCliDS with double GNNs was able to reproduce 36% and 50% of the good clinical decisions (local control and no side effects) and improve 74% and 30% of the bad clinical decisions, respectively. Conclusion: ARCliDS is the first web-based software dedicated to assist KBR-ART with multi-omics data. ARCliDS can learn from the reported clinical decisions and facilitate AI-assisted clinical decision-making for improving the outcomes in DTR.

Intricacies of Human-AI Interaction in Dynamic Decision-Making for Precision Oncology: A Case Study in Response-Adaptive Radiotherapy

“…We found that for NSCLC out of 41 total decision adjustments, 31 (76%) increased and 10 (24%) decreased both TCP and NTCP, whereas for HCC out of 34 total decision adjustments, 9 (26%) increased and 25 (74%) decreased both TCP and NTCP. Note that the simultaneous increase/decrease in TCP and NTCP is by design 22 which was added into the modeling to follow the radiobiological principle that states that increasing/decreasing radiation dose increases/decreases both TCP and NTCP. Since the clinical goal of RT is to maximize TCP while minimizing NTCP, the observation indicates that most of the decisions were adjusted to achieve higher TCP in NSCLC patients and lower NTCP in HCC patients.…”

“…Such AI behavior can be attributed to the training data distribution where the outcome class imbalance was very extreme, i.e. 95 out of 99 patients showed local control ( Table S10 in Niraula et al 22 ). Whereas higher local control is a clinically desirable endpoint and reflects the fact that SBRT is highly successful in treating HCC, it affects the model sensitivity of outcome prediction.…”

“…Adaptive treatment strategies are yet to be widely incorporated in standard clinical practice. Besides KBR-ART, there exist several alternative and complementary adaptive strategies, for instance, adaptation of the number of dose fractionation with fixed dose per fraction, 22,44 during treatment anatomical adaptation via CT-guided 45 or MRI-guided strategies, 46 and gene expression-based adaptation where the full course of radiotherapy is personalized. 47 Thus, evaluators’ level of trust and confidence on AI recommendation would have been higher and more representative if the AI had been designed for standard clinical practice.…”

“…In the response evaluation phase, patients’ treatment response is assessed by comparing pre and during-treatment multi-omics information and in the adaptation phase, a treatment plan is adapted (dose escalation/deescalation). In this study, the clinicians collaborated with ARCliDS 22,25,26 —a software for dynamic decision-making developed using a model-based deep reinforcement learning algorithm. For application in KBR-ART, ARCliDS uses a graphical neural network-based model of radiotherapy environment which defines a patient’s state via a graph of multi-omics features and is capable of assessing treatment response and predicting treatment outcomes.…”

“…We incorporated AIs for two diseases and two treatment modalities and enrolled evaluators from two medical institutions located in different states, with different experience levels, sub-specialties, and professional backgrounds. Considering the fact that evaluators’ mental model of AI can affect human-AI interaction, 32 we took the following steps to homogenize evaluators’ first impression of AI: (i) created two 10-minute tutorial videos; 33,34 (ii) conducted a pre-evaluation information session with the evaluators, in which, we played the training video and demonstrated sample evaluation followed by a question-and-answer round; and (iii) included a web-link to the tutorial video and original ARCliDS manuscript 35 in the evaluation modules. Furthermore, to optimize evaluator’s time utilization, we designed the evaluation to take no more than an hour, added a user account system so the evaluation could be completed in multiple sessions if needed, and enabled automatic saving of all user inputs in the cloud.…”