LBA: Online Learning-Based Assignment of Patients to Medical Professionals

Hanan Rosemarin, Ariel Rosenfeld, Steven Lapp, Sarit Kraus

TL;DR

This paper introduces Learning-Based Assignment (LBA), a novel approach that uses machine learning to dynamically assign patients to medical professionals in real-time. LBA significantly outperforms traditional heuristics in emergency departments and radiology settings by learning institution-specific objectives and adapting to complex, multi-faceted decision-making.

What Problem Does This Paper Address?

The core challenge addressed is the complex and critical task of assigning patients to medical professionals. This isn't just about matching a person to a task; it's about optimizing for multiple, often conflicting, objectives like minimizing patient wait times, ensuring the highest quality of care, and managing limited resources. Current methods often rely on generic heuristics that don't account for the unique needs and priorities of individual medical institutions. This can lead to suboptimal patient outcomes and inefficient resource allocation. The paper focuses on two specific, high-stakes scenarios: assigning patients to physicians in an Emergency Department (ED) and assigning medical scans to radiologists.

Methods

The paper proposes a Learning-Based Assignment (LBA) approach. At its heart, LBA combines offline optimization with machine learning. Here's the breakdown:

  1. Scenario Generation: LBA first creates a large set of realistic, small-scale assignment scenarios based on real-world data and expert input.
  2. Offline Optimization: These scenarios are then solved using optimization techniques (like Mixed Integer Linear Programming or Monte Carlo Tree Search) to find near-optimal solutions. This generates "ground truth" decisions.
  3. Training Data Creation: The optimal decisions from these scenarios are translated into training instances. For each decision made, the system pairs it with all other possible but not chosen decisions.
  4. Learning a Ranking Function: A Deep Neural Network (DNN) is trained as a pairwise ranking function. This function learns to compare two potential assignments and determine which is "better" according to the learned objectives. This is inspired by "Learning to Rank" algorithms used in information retrieval.
  5. Online Policy: The trained DNN then acts as an online policy. When a new event occurs (e.g., a patient arrives, a scan is ready), LBA compares all possible assignments using the trained ranking function and suggests the top-ranked option to the human decision-maker (e.g., a triage nurse or radiologist).

A key aspect is that LBA is designed to be configurable and dynamic, allowing institutions to define their specific objectives and constraints. It also aims to provide a decision support system, not a fully automated one, allowing medical professionals to retain a degree of autonomy.

Datasets

The paper utilizes real-world data from two medical domains:

  • Emergency Department (ED): Data was collected from a major hospital in Israel (Rambam hospital). This data was used to model patient arrival processes, physician availability, and required treatment times. The study specifically focused on the night shift for its complexity.
  • Medical Imaging (Radiology): Data was collected over a three-year period from a major medical institution in Israel. This data included information on studies (scans), readers (radiologists), their subspecialties, processing times, deadlines, and arrival times.

The paper mentions that the code and data used for evaluation are available, which is a significant contribution for reproducibility.

Key Insights and Ideas

  • Tailored Solutions Beat Generic Heuristics: The central idea is that a "one-size-fits-all" approach to patient assignment is insufficient. LBA's ability to learn institution-specific objectives and constraints leads to demonstrably better outcomes.
  • Learning to Rank for Complex Decisions: Applying pairwise ranking via DNNs is a powerful way to model complex, multi-objective decisions where direct scoring might be difficult or less effective. It allows the system to learn subtle preferences.
  • Balancing Autonomy and Optimization: LBA aims to support, not replace, human decision-makers. By providing ranked recommendations, it allows professionals to retain autonomy while still benefiting from data-driven optimization.
  • The "Exposure List" Problem in Radiology: The paper highlights a unique challenge in radiology: managing radiologists' "exposure lists" (the studies they see). LBA offers a way to optimize this exposure to balance reader autonomy with the need to meet study deadlines.
  • Pareto Improvement: In the ED setting, LBA achieved improvements across multiple objectives without negatively impacting others, demonstrating a Pareto improvement over existing methods.

Key Results

The LBA approach demonstrated significant improvements in both tested domains:

  • Emergency Department (ED):
    • Risk of Adverse Consequences: Reduced by 10% (normal flow) and 15% (heavy flow) compared to the FCFSwU heuristic.
  • Wait Times: Slightly reduced by an average of 20 seconds per patient.
  • Length of Stay: Reduced by 5% (normal flow) and 11% (heavy flow).
  • Interruptions: Significantly fewer interruptions (treatment is ~10x less likely to be interrupted with LBA), leading to a ~20x higher penalty cost for FCFSwU.
  • Crowdedness: No significant difference.
  • Medical Imaging (Radiology):
    • LBA significantly outperformed baseline policies (Naïve and H) in terms of overall objective scores, achieving an average score of 69 compared to 82 (H) and 75.5 (Naïve) for a specific configuration.
  • Crucially, LBA was the only algorithm that resulted in zero studies read "extremely late" (more than 60 minutes past deadline).
  • The study also showed that LBA's performance improved with lower exposure list sizes (k), indicating better control and adherence to the system-wide objective.

Limitations and Future Work

  • Training Time: The paper acknowledges that training LBA is more time-consuming than deploying simple heuristics, as it involves scenario generation, offline optimization, and model training. However, this is an offline cost, and the online deployment is fast.
  • Deployment Challenges: Introducing a new automated system may face resistance from medical staff. The authors are actively working on deploying LBA in a real hospital ED.
  • Data Dependency: LBA requires access to real-world data or statistical models. If unavailable, transfer learning might be an option.
  • Feature Engineering: The current approach relies on explicitly defined features. For more complex settings with a vast number of raw features, automatic feature extraction methods will be explored.
  • Future Work:
    • Extending the model to handle assignments involving multiple physicians for a single patient.
  • Applying LBA to other domains, such as task recommendation for crowd-sourcing platforms.
  • Further investigation into the trade-off between reader autonomy (exposure list size) and overall system performance.
  • Exploring more complex allocation objectives and fairness considerations.

Summary updated

Abstract

Central to any medical domain is the challenging patient to medical professional assignment task, aimed at getting the right patient to the right medical professional at the right time. This task is highly complex and involves partially conflicting objectives such as minimizing patient wait-time while providing maximal level of care. To tackle this challenge, medical institutions apply common scheduling heuristics to guide their decisions. These generic heuristics often do not align with the expectations of each specific medical institution. In this article, we propose a novel learning-based online optimization approach we term Learning-Based Assignment (LBA), which provides decision makers with a tailored, data-centered decision support algorithm that facilitates dynamic, institution-specific multi-variate decisions, without altering existing medical workflows. We adapt our generic approach to two medical settings: (1) the assignment of patients to caregivers in an emergency department; and (2) the assignment of medical scans to radiologists. In an extensive empirical evaluation, using real-world data and medical experts’ input from two distinctive medical domains, we show that our proposed approach provides a dynamic, robust and configurable data-driven solution which can significantly improve upon existing medical practices.
BibTeX
@inproceedings{rosemarin2021lba,
  author = {Hanan Rosemarin and Ariel Rosenfeld and Steven Lapp and Sarit Kraus},
  title = {{LBA: Online Learning-Based Assignment of Patients to Medical Professionals}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=7XJB8qaV1X},
}
Figure