Machine Learning for Hospital Capacity and Patient Flow Forecasting

32 concepts28 papers1 proposal

Managing a hospital's capacity is fundamentally a queuing problem under extreme uncertainty. Every day, administrators face a delicate balancing act: if they maintain too many empty beds and idle staff, the hospital wastes expensive, limited resources; if they under-allocate, the system experiences "access block" and severe overcrowding. This overcrowding is not merely an administrative headache—as noted by Karthik Padthe et al. in Emergency Department Optimization and Load Prediction in Hospitals, it is a critical global safety issue directly linked to increased patient mortality, treatment delays, and clinician burnout.

Historically, hospitals have relied on subjective staff heuristics or static historical averages to manage this volatility. However, these methods fail to capture the highly non-linear, dynamic factors that dictate how patients arrive, move through, and exit a healthcare facility.

To solve this, a cohesive research direction has emerged around predictive healthcare modeling. By treating the hospital as an interconnected network, researchers are using machine learning to forecast three critical operational phases: Influx (who is arriving?), Throughput (where are they going and for how long?), and Outflow (when are they leaving?).


Anticipating the Influx via Environmental Proxies

The first step in proactive capacity management is predicting surges before patients arrive at the front door. While emergency arrivals seem random, population-level health is deeply coupled with environmental and behavioral signals. Machine learning models can ingest these non-clinical, public data streams to forecast spikes in volume.

For instance, seasonal weather shifts and air pollution levels are strong physiological triggers for respiratory illnesses. In Peak Outpatient and Emergency Department Visit Forecasting for Patients With Chronic Respiratory Diseases Using Machine Learning Methods: Retrospective Cohort Study, Junfeng Peng et al. demonstrated that combining meteorological data with air quality indices allows machine learning models to accurately predict "peak demand days" for chronic respiratory conditions. Similarly, Grace Guan and Barbara Engelhardt utilized recurrent neural networks (RNNs) in Predicting Sick Patient Volume in a Pediatric Outpatient Setting using Time Series Analysis to capture complex weekly and seasonal fluctuations, significantly outperforming traditional linear baselines that mistakenly assume a constant daily volume.

This environmental coupling extends to acute cardiovascular events as well. In the study Machine learning-based forecasting of daily acute ischemic stroke admissions using weather data, the authors showed that integrating local meteorological data into forecasting models reduces prediction errors to within approximately 1.2 stroke cases per day, allowing facilities to dynamically adjust specialized staffing.

Even when direct clinical data is entirely unavailable due to privacy constraints, proxy data can fill the gap. Haipeng Chen et al. proved this in Using Public Data to Predict Demand for Mobile Health Clinics by using public bike-share usage and weather forecasts as proxies for foot traffic, reducing clinic demand prediction errors by up to 51.8% compared to historical averages.


Modeling Internal Throughput and Patient Pathways

Once a patient enters the hospital, the next challenge is predicting their trajectory. A common pitfall in hospital planning is treating a patient's Length of Stay (LOS) as a single, static number. In reality, a patient's journey is a dynamic sequence of transfers between different units (e.g., Emergency Department $\rightarrow$ Intensive Care Unit $\rightarrow$ General Ward).

In Importance of patient bed pathways and length of stay differences in predicting COVID-19 hospital bed occupancy in England, Quentin J. Leclerc et al. illustrated why granular modeling matters: relying on simple average lengths of stay consistently underestimates actual bed occupancy during public health crises. This happens because averages wash out the extreme "right-tail" delays caused by complex multi-stage transfers. By modeling explicit "bed pathways," administrators can anticipate compounding internal bottlenecks.

To make these pathway models actionable, hospitals must predict a patient's disposition and resource needs early in their visit. Long Song et al. addressed this in Developing interpretable machine learning models to predict length of stay and disposition decision for adult patients in emergency departments by utilizing transparent Lasso models. This approach achieves high predictive accuracy for ED admission decisions while ensuring clinicians can see exactly which clinical factors are driving the prediction.

Furthermore, early triage can isolate patients who do not require long-term resources. Michael Dinh et al. developed a clinical prediction rule in Deriving a prediction rule for short stay admission in trauma patients admitted at a major trauma centre in Australia to identify trauma patients likely to require a "short stay" ($\le$ 48 hours). Identifying these patients early allows hospitals to route them to dedicated short-stay units, preserving main ward beds for more critical cases.


Unlocking the Bottleneck via Outflow Forecasting

Emergency Department overcrowding is rarely just an influx problem; it is frequently an "access block" caused by a lack of open inpatient beds upstairs. If the back door of the hospital is blocked, the front door stalls. Therefore, predicting discharges is just as vital as predicting admissions.

A major hurdle is that real-time clinical data can be siloed or delayed. To circumvent this, Shivapratap Gopakumar and Dinh Q. Phung demonstrated in Forecasting Daily Patient Outflow From a Ward Having No Real-Time Clinical Data that machine learning models (such as Random Forests) can accurately forecast daily patient outflow using only basic, readily available administrative data.

When real-time data is accessible, it can be used to shift hospital operations from a reactive stance to a proactive one. In Improving patient flow during infectious disease outbreaks using machine learning for real-time prediction of patient readiness for discharge, Hamza Javed developed a framework that continuously ranks hospitalized patients by their "discharge readiness" within the next 24 hours. This provides a real-time decision-support tool that allows ward managers to coordinate with the ED ahead of time, ensuring that as soon as a patient is ready to leave, a bed is cleared for the next admission.


System-Wide Optimization and Advanced Architectures

As these predictive components mature, researchers are scaling them from individual wards to entire regional healthcare systems by combining advanced deep learning architectures with mathematical optimization.

On the macro-forecasting side, standard time-series models often struggle with regional variations. To capture these nuances, Shihao Yang et al. designed the County Augmented Transformer (CAT) in County augmented transformer for COVID-19 state hospitalizations prediction. CAT improves state-level hospitalization forecasts by explicitly augmenting macro trends with granular, county-level information through a specialized transformer architecture. Similarly, Steffen Albrecht et al. utilized advanced time-series foundation models, such as the Temporal Fusion Transformer (TFT), in Forecasting severe respiratory disease hospitalizations using machine learning algorithms to successfully navigate the high year-to-year variability of viral circulation.

[External Signals] (Weather, Air Quality, Public Proxies)
       │
       ▼
[Influx Forecasting] ──► [Throughput & Pathway Modeling] ──► [Outflow Forecasting]
       │                            │                                 │
       └────────────────────────────┴─────────────────────────────────┘
                                    │
                                    ▼
                     [System-Wide Optimization] (e.g., Column Generation)

Ultimately, these predictions serve as the foundation for automated, system-wide resource allocation. In Radiation therapy patient scheduling using a column generation approach: an algorithmic framework and a single-institution study, Sara Frimodig et al. demonstrated how predictive insights can be directly integrated into operational scheduling. By using a Column Generation optimization framework that models both scheduled appointments and uncertain, urgent patient arrivals, their algorithmic approach reduced average patient waiting times by 80%.

By moving away from static averages and embracing dynamic, pathway-aware machine learning models, modern healthcare systems can transition from a state of constant crisis management to a proactive, optimized flow of patient care.

Key Foundational Work

Forecasting Daily Patient Outflow From a Ward Having No Real-Time Clinical Data (2016) — contributed 1 concept elements to this direction.
Deriving a prediction rule for short stay admission in trauma patients admitted at a major trauma centre in Australia (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_emergency_department_overcrowding_37da39de,
  title = {emergency department overcrowding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/emergency-department-overcrowding/rdr_37da39dedcb648cb97a50f5913d2370a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_emergency_department_overcrowding_37da39de_v1,
  title = {emergency department overcrowding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/emergency-department-overcrowding/rdr_37da39dedcb648cb97a50f5913d2370a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake