Machine learning to predict curative multidisciplinary team treatment decisions in oesophageal cancer.

BACKGROUND: Rising workflow pressures within the oesophageal cancer (OC) multidisciplinary team (MDT) can lead to variability in decision-making, and health inequality. Machine learning (ML) offers a potential automated data-driven approach to address inconsistency and standardize care. The aim of this experimental pilot study was to develop ML models able to predict curative OC MDT treatment decisions and determine the relative importance of underlying decision-critical variables. METHODS: Retrospective complete-case analysis of oesophagectomy patients ± neoadjuvant chemotherapy (NACT) or chemoradiotherapy (NACRT) between 2010 and 2020. Established ML algorithms (Multinomial Logistic regression (MLR), Random Forests (RF), Extreme Gradient Boosting (XGB)) and Decision Tree (DT) were used to train models predicting OC MDT treatment decisions: surgery (S), NACT + S or NACRT + S. Performance metrics included Area Under the Curve (AUC), Accuracy, Kappa, LogLoss, F1 and Precision -Recall AUC. Variable importance was calculated for each model. RESULTS: We identified 399 cases with a male-to-female ratio of 3.6:1 and median age of 66.1yrs (range 32-83). MLR outperformed RF, XGB and DT across performance metrics (mean AUC of 0.793 [±0.045] vs 0.757 [±0.068], 0.740 [±0.042], and 0.709 [±0.021] respectively). Variable importance analysis identified age as a major factor in the decision to offer surgery alone or NACT + S across models (p < 0.05). CONCLUSIONS: ML techniques can use limited feature-sets to predict curative UGI MDT treatment decisions. Explainable Artificial Intelligence methods provide insight into decision-critical variables, highlighting underlying subconscious biases in cancer care decision-making. Such models may allow prioritization of caseload, improve efficiency, and offer data-driven decision-assistance to MDTs in the future.

Trustworthy human-AI partnerships.

In this paper, we foreground some of the key research challenges that arise in the design of trustworthy human-AI partnerships. In particular, we focus on the challenges in designing human-AI partnerships that need to be addressed to help humans and organizations trust their machine counterparts individually or as a collective (e.g., as robot teams or groups of software agents). We also aim to identify the risks associated with human-AI partnerships and therefore determine the associated measures to mitigate these risks. By so doing, we will trigger new avenues of research that will address the key barriers to the adoption of AI-based systems more widely in our daily lives and in industry.

Decentralized Patrolling Under Constraints in Dynamic Environments.

We investigate a decentralized patrolling problem for dynamic environments where information is distributed alongside threats. In this problem, agents obtain information at a location, but may suffer attacks from the threat at that location. In a decentralized fashion, each agent patrols in a designated area of the environment and interacts with a limited number of agents. Therefore, the goal of these agents is to coordinate to gather as much information as possible while limiting the damage incurred. Hence, we model this class of problem as a transition-decoupled partially observable Markov decision process with health constraints. Furthermore, we propose scalable decentralized online algorithms based on Monte Carlo tree search and a factored belief vector. We empirically evaluate our algorithms on decentralized patrolling problems and benchmark them against the state-of-the-art online planning solver. The results show that our approach outperforms the state-of-the-art by more than 56% for six agents patrolling problems and can scale up to 24 agents in reasonable time.

Multi-Agent Patrolling under Uncertainty and Threats.

We investigate a multi-agent patrolling problem where information is distributed alongside threats in environments with uncertainties. Specifically, the information and threat at each location are independently modelled as multi-state Markov chains, whose states are not observed until the location is visited by an agent. While agents will obtain information at a location, they may also suffer damage from the threat at that location. Therefore, the goal of the agents is to gather as much information as possible while mitigating the damage incurred. To address this challenge, we formulate the single-agent patrolling problem as a Partially Observable Markov Decision Process (POMDP) and propose a computationally efficient algorithm to solve this model. Building upon this, to compute patrols for multiple agents, the single-agent algorithm is extended for each agent with the aim of maximising its marginal contribution to the team. We empirically evaluate our algorithm on problems of multi-agent patrolling and show that it outperforms a baseline algorithm up to 44% for 10 agents and by 21% for 15 agents in large domains.