TLDR: Feasibility-Guided Fair Adaptive Reinforcement Learning (FG-FARL) is an offline reinforcement learning method for Medicaid population health management. It calibrates per-group safety thresholds to reduce harm and equalize fairness targets (coverage or harm) across protected subgroups. Evaluated on de-identified Medicaid data, FG-FARL achieves comparable value to baselines while significantly improving fairness metrics, offering a practical and auditable path to safer and more equitable AI-driven decision support in healthcare.
This research introduces an innovative approach called Feasibility-Guided Fair Adaptive Reinforcement Learning (FG-FARL), designed to enhance decision support in Medicaid population health management. The core idea is to use artificial intelligence (AI) to recommend actions for care coordination while ensuring both safety and fairness for all individuals.
Medicaid programs serve people with complex health needs, and care teams often face the challenge of deciding who to contact, how to reach them (e.g., phone, home visit), and when to follow up. These decisions are critical, as resources are limited, and inappropriate actions can inadvertently cause harm or increase burden on patients. FG-FARL aims to provide AI-driven recommendations that improve outcomes, avoid harm, and ensure equitable access to care and risk of harm across different demographic groups like age, sex, or race.
FG-FARL works by first learning a model that predicts the risk of harm for each patient’s state. Based on this risk model, it sets specific safety thresholds for different patient subgroups. These thresholds can be adjusted in two ways: either to ensure an equal proportion of “safe” states for each group (coverage mode) or to cap the observed harm rate within the safe actions for each group (harm mode). This means the system actively works to prevent any single group from being disproportionately exposed to risk or denied access to beneficial interventions.
After establishing these group-specific safety thresholds, FG-FARL then learns a “preference policy.” This policy recommends the best next actions (like who to contact, the outreach method, or follow-up timing) only from the set of actions deemed safe for that particular patient and their group. This two-step process makes the system transparent and auditable, allowing healthcare providers to understand and adjust the safety and fairness parameters.
The researchers evaluated FG-FARL using de-identified patient data from a Medicaid population health management program. They compared it against existing methods like behavior cloning (BC) and HACO (a global safety baseline). The results showed that FG-FARL achieved similar overall effectiveness in improving patient outcomes compared to the baselines, but significantly improved fairness metrics. This demonstrates that it’s possible to create AI decision support systems that are both effective and equitable in real-world healthcare settings.
The actions considered by the system include various care coordination activities such as low-intensity outreach, scheduling assistance, social service referrals, clinician coordination, and high-intensity outreach like home visits. The system uses patient states, including demographics and utilization indicators, and defines rewards based on adverse events, aiming to minimize harm.
A key aspect of FG-FARL is its interpretability. The safety and fairness “dials” (parameters like alpha and epsilon) are explicit, making it easier for clinicians and care teams to understand and trust the recommendations. The preference policy itself is a straightforward classifier, further enhancing transparency. The system is also computationally efficient, completing its training and auditing in minutes, making it practical for routine use in healthcare operations.
Also Read:
- Enhancing Medical AI: A Causal Approach to Handling Incomplete Multimodal Data
- Reinforcement Learning Unlocks Advanced Reasoning in Large Language Models
This research offers a practical pathway for health plans and provider organizations to deploy AI systems that support care coordination, ensuring that decisions are not only effective but also fair and safe for all members of the Medicaid population. For more technical details, the full research paper can be accessed here: Feasibility-Guided Fair Adaptive Offline Reinforcement Learning for Medicaid Population Health Management.
