This project aims to develop a recommender system for referral of specialist care doctors to be used by primary care doctors, that is compatible with current referral practice, and that can transparently encourage organizational change, towards a more effective patient-centric healthcare management. Typically, for a given patient with clinical needs, the primary care physicians can make a choice of several specialists to whom they may refer, and their choice would have important downstream effects. As such, primary-specialty referral may affect many aspects of patient care, such as quality of care, patient satisfaction and healthcare costs, etc. Researchers recently leveraged the patient consultation history extracted from insurance claims data to construct the patient sharing network between physicians based on the shared patients. Essentially, the patient sharing network operationalizes an informal information-sharing network in which physicians provide care to shared patients. This network does not necessarily conform to the formal organizational structure that physicians are affiliated with but may provide valuable insights in explaining the referral mechanism. For example, social network analysis on Medicare administrative data showed that the structure of patient sharing networks and the position of physicians in the network has a significant relationship with the overall cost and intensity of care. These metrics derived from network science can serve as informative features to boost predictive model performance and optimize the health system for improved medical outcomes. This project will analyse the current state of medical referral in CUF Health and compare it with another reality in the USA. We will implement novel recommendation systems that are: 1) trustworthy, 2) advised by policy, 3) not disruptive of the usual referral process used today, 4) fostering online communities.

Primary care serves as patients’ first point of contact with the healthcare system and is a continuing focal point of comprehensive, accessible, and community-based care. They also perform a gate-keeping process for specialist referrals. Typically, the patients will receive several specialist doctor recommendations, and their choices could significantly affect their own health outcomes. We will build upon our previous work with two CUF Health datasets covering patient transaction data and Human Resources pseudo-anonymized database covering the years 2013 - 2017 to update our analysis and identify possible drifts in patient, primary care, and specialist doctors behavior, and analyse it differentially with US data.

Simultaneously, we will develop prediction models that can transparently incorporate policy incentives, and we will explore with CUF Health and doctors the exact formulation of the needed policy changes. In the second half of the year we will apply the developed methodologies and test them in terms of transparency, accuracy, and effectiveness regarding the intended policy changes. All our work will be closely monitored by the ethical committees of the institutions involved. The ethics assessment will relate to data usage and to ethical use of AI in healthcare. In order to comply with the highest standards of ethical and trustworthy AI we will implement transparent, white-box methods, benchmarked against black-box models. In this project, we hypothesize that medical doctors' informal social networks can influence the referral process. Doctors' referral decisions may be limited to their social contacts which will not always benefit patient-centered care. This implies that network structure metrics derived from the doctors' social network can serve as informative features to boost the predictive performance of a model for referral recommendations. As such, we plan to create two networks: 1) the referral network connecting primary care physicians to specialist physicians if a patient consults a primary care physician and then a specialist physician within a month, and 2) the social network of all doctors according to their similar social and educational profiles. Then, we learn the representation of the referral network using a Graph Neural Network (GNN). The main objective of the work is to uncover hidden mechanisms in the primary-specialty referrals using features extracted from the informal social network of doctors, which may help health organizations to improve the referral process through recommendations. The referral mechanisms encoded in the learned representation can contain both relevant and circumstantial factors. A relevant factor might be the particularly suited skills of the specialist physician in the specific clinical context of the patient, as identified by the primary care physicians. A circumstantial factor can be that the primary care physicians cannot remember all suitable specialist physicians within the provider’s network, and her selection is suboptimal. The ML model trained on historical data will capture referral patterns and its predictive power will enclose both types of factors. It is crucial for a useful Healthcare Recommender System to bring clarity to those factors, and to capture only the relevant ones. For this clarity to emerge, we will study the behavior of our learned referral model and use structural causal learning and causal reinforcement learning to explain the reasoning behind each referral decision. The approach is motivated by known limitations of additive explanation models currently used, e.g., LIME or SHAP, that are not sensitive to network structure and cannot perceive complex, non-additive influences. The outcomes of this development will produce a first version of our healthcare recommender system. A second aim is to develop a fully interpretable model without the need to train a causal model to learn the predictive model. This is an important step towards transparency and trustworthiness in the system among doctors and patients because those stakeholders will be fully aware of the criteria behind the recommendation. These two developments will translate forefront research in ML into the practical realm. The research team has experience both in causal inference and in fully transparent ML models, and with similar datasets, thus the application to the present datasets will not offer major difficulties.

This project is interdisciplinary, as it delivers in healthcare management, business and ML. It will impact machine learning research, as we will develop (1) a novel black-box model explanation framework, leveraging on causal inference and causal reinforcement learning; (2) a methodology for training and inference of ML models, where relevant factors are treated as signal and circumstantial factors as noise, to be disregarded by the learned model. These developments are supported by causal and counterfactual inference and statistical techniques like blind source separation. The project will also impact healthcare management and delivery, as more adequate doctors will be recommended to each patient. Although the models will be trained in CUF data, we will test performance in data from, e.g., the USA. Risk assessment and mitigation. In growing magnitude, the main risks are: (1) not enough data availability; (2) predictive models not achieving good enough performance; (3) causal explanations not coherent with domain knowledge; (4) separation of relevant and circumstantial factors not improving the HRS human-evaluated performance. On (1): From our ongoing collaboration with CUF on two research topics, data accessibility is not a problem. On (2): We have already worked with a similar CUF dataset in the past and we trained performant models on these data. On (3): Causal models will be kept simple enough to allow experts to understand the model explanations. We will maintain open communication with stakeholders at CUF to help define explanation concepts. On (4): This point is the high-risk high-gain aspect of this project. The research team is well-equipped to address it, with experience in network science, information theory and statistical learning. Even if the project fails to deliver this development, we are still providing relevant datasets, data analysis, and predictive and explanation ML models that are a contribution in healthcare management and ML research.