REVIEW 3 major objections 4 minor 17 references
Using hourly aggregate bed requests driven by admission predictions can cut average boarding time for admitted ED patients by 30–70 percent and overall ED length of stay by 6–15 percent, at modest idle-bed cost.
Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →
Prediction-driven aggregate early bed requests reduce simulated ED boarding time by 30–70% and ED length of stay by 6–15% with modest prepared-bed idle time.
T0 review reviewed 2026-08-01 challenge →
load-bearing objection Solid system-level early bed-request framework with careful simulation; headline 30–70% boarding reduction is conditional on borrowed hospital-side parameters and needs sensitivity analysis before being taken at face value. the 3 major comments →
Proactive Inpatient Bed Requests for Emergency Department Admissions
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
Core claim
The central discovery, on the paper's own terms, is that the timing and quantity of inpatient bed requests can be treated as an ED-level inventory problem instead of a per-patient decision. The paper's MDP state tracks patients by admission-probability class, the number boarding (or idle beds, signed), and the number of requested beds still being prepared. The action is how many additional beds to request. From this formulation it derives an MDP-based heuristic, a newsvendor heuristic that matches the mean and variance of future demand to a triangular distribution over an eight-hour horizon, and a deep-Q-learning heuristic. In a discrete-event simulation of a 2019 ED, the newsvendor policy p
What carries the argument
The load-bearing object is the aggregate hourly request quantity, computed from a scalar imbalance between expected near-term admissions and bed supply—the paper writes it as φ = Σ α_i n_i + k − r, where α_i is a patient's admission probability, n_i counts patients in that class, k is the number boarding (negative if beds are idle), and r is the number of requested beds still in preparation. Three policies use this state in different ways: MHP applies a linear policy from an auxiliary finite-horizon MDP; NHP solves a newsvendor problem whose demand distribution is a symmetric triangular fit to the mean and variance of admissions-minus-ready-beds over eight hours; DQN learns a Q-function over
Load-bearing premise
The 30–70% boarding reduction is produced by a simulation in which hospital-side inputs—transfer-preparation time, inpatient length-of-stay distribution, and the number of beds reserved for ED admissions—are borrowed from other hospitals or tuned to match observed ED data, not estimated from the study ED's own records; different hospital dynamics could shrink the benefit.
What would settle it
Measure transfer-preparation times, inpatient lengths of stay, and reserved-bed capacity from a single hospital, re-run the same policy comparison in that hospital's own simulation, and check whether the 30–70% boarding reduction and 1–95 minute idle-time range still hold. Alternatively, run a six-month pilot using hourly aggregate newsvendor requests on a general-medicine unit and compare boarding times with the prior year; a reduction well below 30% or idle times far above 95 minutes would refute the quantitative claim.
If this is right
- An ED with a two-class triage admission predictor (low admission probability around 8.5%, high around 57.7%) can expect boarding time to drop from roughly 248 minutes to 77–179 minutes, depending on how much idle bed time the hospital tolerates.
- Total ED length of stay for all patients drops by about 26–65 minutes (6–16%), which can shorten waits for other ED patients by freeing beds sooner.
- Even a simple greedy policy—requesting the expected number of admissions minus beds already requested—captures a large share of the benefit, but without tunable cost parameters.
- The newsvendor heuristic is the best default for the boarding/idle tradeoff, while the reinforcement-learning heuristic is preferable when downstream hospital stability depends on a smooth bed-request process.
- Benefits persist with only 70% of beds interchangeable and under a six-week pandemic surge with expanded hospital capacity, so the policy need not wait for perfectly general bed pools.
Where Pith is reading between the lines
- Beyond the paper, the aggregate-request logic suggests a low-tech implementation: hospital managers could start with the newsvendor formula and an eight-hour lookahead, requiring only admission probabilities and a rough bed-preparation-time mean.
- An untested corollary is that the value of proactive requests should increase with bed-preparation time: the longer the transfer-preparation process, the more boarding is avoided by starting it early. The paper does not report this comparative static.
- The quantitative claims are simulation-based; the natural next step, which the paper explicitly flags, is a controlled pilot in a real ED. If the pilot uses only general-medicine beds, the 70% eligibility result is the relevant benchmark.
- Because the MDP treats beds as interchangeable, the framework's benefits may concentrate in hospitals with large general-medicine admitting units; specialized hospitals would likely need a hybrid of aggregate and per-patient requests.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes an MDP-based framework for generating aggregate early inpatient bed requests from an emergency department, using admission-probability and time-to-disposition predictions for current ED patients. Three heuristics are developed: an approximate dynamic programming policy (MHP), a newsvendor-type policy (NHP), and a deep Q-learning policy (DQN). The policies are evaluated in a high-fidelity Arena simulator calibrated to 2019 data from a large North Carolina ED. The central quantitative claim is that proactive aggregate bed requests reduce average boarding time for admitted patients by 30–70% and average ED length of stay by 6–15%, at the cost of modest average idle time for prepared inpatient beds. Analytical results (Propositions 4.1–4.5) are stated with proofs in the online supplement, and the simulation reporting is detailed: 1,000-year batch-means runs, 95% CI half-widths, and full parameter tables for all scenarios.
Significance. If the simulation results are accepted, the paper would make a useful operational contribution: it shows that a relatively simple, interpretable newsvendor heuristic can capture most of the benefit of more complex RL-based approaches, with an explicit trade-off between ED performance and inpatient-bed idle time. The analytical framework is general, and the decision-tree approximation of MHP/DQN is a pragmatic step toward implementation. The paper is also transparent about data limitations and uses a plausible dual-simulator architecture. However, the headline quantitative claims rest on hospital-side model components (transfer preparation time, hospital length of stay, and reserved bed capacity) that are not estimated from the study hospital's own data; the paper even states that its data 'lack the granularity' to estimate them (OS.2.2). This is a significant caveat because the CP baseline itself is an output of those assumptions, and the proportional reductions are therefore predictions of a model whose counterfactual validity has not been tested. The central mechanism is credible and the policy comparisons are internally consistent, but the quantitative ranges should be regarded
major comments (3)
- [OS.2.2–OS.2.3, Section 5.1] The hospital-side parameters that drive the headline reductions are not estimated from the study ED's 2019 data. Transfer preparation time is a truncated lognormal taken from Shi et al. (2016) (mean 3.3h, CV 0.6, max 12h); hospital LoS is adapted from Shi et al. (2014) (a Singaporean hospital); and the reserved bed capacity of 337 is a tuning parameter selected to match observed second-stage durations (OS.2.3). Since boarding time under CP and under every proactive policy is co-determined by TPP, hLoS, and bed capacity, the claimed 30–70% boarding reduction and 6–15% LoS reduction are conditional on these unvalidated inputs. A sensitivity analysis over TPP mean/CV, hLoS distribution, and reserved bed count (e.g., ±20–30%) is necessary to establish the robustness of the headline ranges. Without it, the abstract's quantitative claims overstate what has actually been demonstrated.
- [Section 5, 'Approximation by decision trees'] The MHP and DQN policies are approximated by decision trees with 150–300 nodes, and Tables 13–16 report the performance of these tree-approximated policies. No fidelity metric is provided: the paper does not report the fraction of state-action pairs where the tree disagrees with the original heuristic, nor the performance difference between the tree policy and the original policy. If the decision-tree approximation is lossy, the relative performance of MHP and DQN in the Arena evaluation could be materially degraded, and the recommendation to prefer NHP (or the claim that DQN is smoother) might be an artifact of the approximation. Please report the approximation error or validate the tree policies against the original policies in a subset of scenarios.
- [Section 5.2 (SOS-P)] Partial bed eligibility is modeled by multiplying the heuristic's request quantity by 0.7. This implicitly assumes that the decision maker knows the exact eligibility fraction and that the demand from ineligible patients can be correctly handled by simple scaling. A more faithful implementation would restrict early requests to eligible patients or incorporate eligibility into the state and action space. As written, the SOS-P results are an approximate sensitivity check rather than a validation of the policies under the stated 'partial bed requests' condition; the abstract's claim that benefits persist in that setting is therefore weaker than presented.
minor comments (4)
- [Section 5, OS.2.1] Typo: 'Phyton simulator' should be 'Python simulator'.
- [Abstract vs. Section 5.1] The abstract states LoS reduction of 6–15%, but Section 5.1 says 'by 6–16%' and Table 13 shows reductions up to about 16%. Please reconcile the range.
- [Table 13] Header typo: 'T uning Parameters' should be 'Tuning Parameters'.
- [Section 4.3, Proposition 4.5] The NHP uses a symmetric triangular approximation for the demand distribution, justified by closed-form tractability. Since NHP is a recommended policy, it would strengthen the paper to report the sensitivity of NHP's trade-off frontier to the demand-shape assumption (e.g., a normal or empirical distribution), even if only in an appendix.
Circularity Check
No load-bearing circularity; the headline reductions are simulation outputs, while the main caveat is borrowed/tuned hospital-side parameters, which is a validity risk rather than a circular reduction.
full rationale
The paper's claimed reductions in boarding time and ED length of stay are outputs of an Arena simulator, not identities, fitted values, or self-citation chains. The theoretical derivations are self-contained: Proposition 4.1 follows from standard MDP results; Proposition 4.2 is proven from the quadratic cost and linear transition structure of the auxiliary MDP; Proposition 4.3 and Corollary 4.1 are derived from convex optimization of the newsvendor objective. The newsvendor demand distribution is estimated from state-dependent mean and variance formulas and then used in a closed-form request quantity; the resulting policy's performance is then measured in a separate, higher-fidelity Arena simulator. No equation equates the predicted boarding reduction or LoS reduction to a tuned parameter or to the benchmarks by construction. The Arena simulator is described as 'a modified version of the one used in Chen et al. [2023]', a self-citation with overlapping authors, but the current paper redoes the input analysis using 2019 data and validates the simulator against observed ED LoS and first- and second-stage durations, so the cited simulator is not an unverified self-citation that carries the argument alone. The authors explicitly flag the main model risks: hLoS parameters are 'adapt[ed] from Shi et al. [2014]' because the data 'lack the granularity' (OS.2.2), TPP is borrowed from Shi et al. [2016] as a truncated lognormal with mean 3.3 hours, and the reserved hospital bed count of 337 is 'a tuning parameter' selected to align simulated second-stage durations with observed data (OS.2.3). These are calibration and external-validity concerns for the counterfactual, not circularity: the CP baseline itself is a simulator output, and the proposed policies are separately evaluated in the same simulator rather than computed from the fitted inputs. The paper could be strengthened by a sensitivity analysis over TPP and reserved capacity, but the derivation chain does not reduce its predictions to its own inputs.
Axiom & Free-Parameter Ledger
free parameters (7)
- Reserved hospital bed capacity =
337 beds
- Early bed-request cap =
55 outstanding requests (65 under PRS)
- MHP cost-to-go weights (α̃, β̃) =
e.g., (1.4,18), (1.6,12), (0.5,5), (0.5,2), (0.5,0)
- NHP cost parameters (êb, êc) and horizon ψ =
e.g., (256,1) to (10,128) with ψ=8
- DQN cost weights (cb, ce) =
C1={(10,1),(4,1),(1,1),(1,4),(1,10)}; C2 analogues
- Type-I/II admission probabilities =
0.0848 and 0.577
- Transfer preparation time (TPP) parameters =
truncated lognormal, mean 3.3h, CV 0.6, max 12h
axioms (6)
- standard math Existence/uniqueness of the optimality-equation solution via Puterman's Theorem 6.10.4 (infinite-horizon discounted MDP, countable state space)
- domain assumption Infinite ED and hospital bed capacity in the analytical MDP
- domain assumption Hospital beds are interchangeable across admitted patients
- domain assumption Class-constant disposition probabilities and exponential service/bed-preparation times in the Python simulator
- ad hoc to paper Bed-request demand is approximated by a symmetric triangular distribution with matched mean and variance
- domain assumption Hospital length-of-stay and transfer-preparation parameters transfer from Shi et al. 2014 (Singapore) and Shi et al. 2016
Cite this review
Pith. "Pith review of Proactive Inpatient Bed Requests for Emergency Department Admissions." pith.science (2026). https://pith.science/paper/RLNDASVW
@misc{pith2026260715432,
author = {Pith},
title = {Pith review of: Proactive Inpatient Bed Requests for Emergency Department Admissions},
year = {2026},
howpublished = {\url{https://pith.science/paper/RLNDASVW}},
note = {Machine review of arXiv:2607.15432}
}
read the original abstract
Emergency department (ED) boarding occurs when admitted patients remain in the ED while awaiting inpatient beds. Boarding is a major driver of ED crowding and has been associated with poor patient outcomes. We propose a framework to help EDs reduce boarding time and length of stay by using information about current patients and bed availability to proactively request inpatient beds before admission decisions are finalized. We formulate the problem as a Markov decision process in which predictions of each patient's admission probability and time to disposition are aggregated to guide early inpatient bed requests. This formulation leads to three data-driven policies based on approximate dynamic programming, reinforcement learning, and a newsvendor-type approach. Using a simulation model based on data from a large ED, we evaluate these policies across a wide range of settings. The simulation study shows that proactive aggregate bed requests can reduce average boarding times for admitted patients by 30-70\% and average length of stay for all ED patients by 6-15\%, while creating only modest idle time for prepared inpatient beds. The newsvendor heuristic provides the most attractive tradeoff between ED performance and inpatient bed idle time, whereas the reinforcement learning heuristic produces smoother bed-request patterns when stability in downstream hospital processes is especially important. Our work shows how EDs can use prediction tools to make proactive bed-request decisions that improve ED operations while helping managers balance reductions in ED delays against inpatient bed idle time. Our findings also illustrate the value of evaluating both simple myopic heuristics and more sophisticated reinforcement learning-based approaches, since each can offer distinct advantages depending on the performance measures and implementation constraints most important to managers.
Figures
Reference graph
Works this paper leans on
-
[3]
T able 3:Suitable ED Care Areas with respect to Age Group, Triage Level, and Priority Order
Table 3 summarizes these classifications. T able 3:Suitable ED Care Areas with respect to Age Group, Triage Level, and Priority Order. Age Group ESI Priority Order Suitable Care Areas (In order of preference as capacity allows) Adult 1 1 Team A Reserved, Team A, Team B Adult 2 2 Team A, Team B Adult 3 (Acute) 3 Team A, Team B Adult 3 (Non-acute) 4 Team A,...
2016
-
[4]
Technical report, Agency for Healthcare Research and Quality,
Implementation Handbook 2012 Edition. Technical report, Agency for Healthcare Research and Quality,
2012
-
[7]
com/en-us/products/software/arena-simulation.html
URLhttps://www.rockwellautomation. com/en-us/products/software/arena-simulation.html. Accessed: 2026-06-22. Trevor Hastie, Robert Tibshirani, and Jerome Friedman.The Elements of Statistical Learning: Data Mining, Inference, and Prediction. New York: Springer, second edition,
2026
-
[10]
A2 +c max MX m=1 nm +N+|k|+|r|+a !# ≤ X j∈S,N≥0 p(j, N|s, a)
hold, then there exists a unique solution to the optimality equation, and if there exists a policyπ ∗ that chooses the action that minimizes the right-hand side of optimality equation (2), thenπ ∗ is the optimal policy that is associated with optimal value functionV(s). To prove that Assumptions 6.10.1 and 6.10.2 hold, suppose the current state is denoted...
2014
-
[11]
When−δ−ξ+ξ √2ece/√ece +ecb < kt ≤(ecb −ece)/4−δ,eymust be in (δ, δ+ξ] for Equation (6) to have a solution
For all other values ofk t, we then find the solutioneyto Equation (6) and lety ∗ =ey+k t. When−δ−ξ+ξ √2ece/√ece +ecb < kt ≤(ecb −ece)/4−δ,eymust be in (δ, δ+ξ] for Equation (6) to have a solution. In this case, solving the corresponding equation 1− (δ+ξ−y) 2 2ξ2 + 2y ecb +ece = ecb −2k t ecb +ece fory, we obtain ey=δ+ξ+ 2ξ2 −ξ p 4ξ2 + 2(ece +ecb)(ece + 2...
2019
-
[12]
The estimated admission probabilities are 0.0848 for Type-I and 0.5770 for Type-II patients
After completing ED service, a patient is either discharged or admitted according to a Bernoulli disposition decision. The estimated admission probabilities are 0.0848 for Type-I and 0.5770 for Type-II patients. Discharged patients leave the system immediately, whereas admitted patients 6 board until a previously requested hospital bed becomes available a...
2023
-
[14]
[2023], TPP follows a truncated lognormal distribution with a mean of 3.3 hours, coefficient of variation of 0.6, and maximum of 12 hours
and Chen et al. [2023], TPP follows a truncated lognormal distribution with a mean of 3.3 hours, coefficient of variation of 0.6, and maximum of 12 hours. Hospital Stay:Since our data covers only ED encounters, we lack direct estimates for hospital length of stay (hLoS). Furthermore, it is not straightforward to set a capacity to hospital beds used by ED ...
2023
-
[15]
For hLoS-day, we primarily use parameters from Table 11 of Shi et al
with minor modifications. For hLoS-day, we primarily use parameters from Table 11 of Shi et al. [2014], distinguishing between patients admitted before noon (ED-AM) and after noon (ED-PM) as in Shi et al. [2016]. According to our partner hospital’s operational practices, we consider that an ED-AM patient is a same-day patient (discharged on the same day a...
2014
-
[17]
We raise the cap for the outstanding early bed requests from 55 to 65 under PRS. OS.2.5 Input Distributions for the Arena Simulator In this section, we provide a summary of the probability distributions and parameter estimates that are implemented in the Arena simulator. Tables 5 through 10 provide the arrival rates estimated from the 2019 arrival data an...
2019
-
[337]
The simulation runs for 1,000 years with a one-year warm-up
Figure 7 compares average ED LoS, first-stage duration, and second-stage duration on an hourly basis using both observed and simulated data. The simulation runs for 1,000 years with a one-year warm-up. On thex-axis, each value denotes a one-hour slot (e.g., 0 is midnight to 1:00 am, and 1 represents 1:00 am to 2:00 am). ED LoS averages are computed over p...
2019
-
[2014]
http://dx.doi.org/10.2139/ssrn.2517050. 32 Online Supplement OS.1. Proofs of Theoretical Results Proof of Proposition 4.1:This proof uses theorems and propositions from Puterman [2014]. According to Theorem 6.10.4 of Puterman [2014], since the state space is countable for our problem, if Assumptions 6.10.1 and 6.10.2 of Puterman
-
[2015]
Deep reinforcement learning with double Q-learning
Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double Q-learning. InProceedings of the Thirtieth AAAI Conference on Artificial Intelligence, page 2094–2100,
2094
-
[2016]
31 Qihao Wu, Jiangxue Han, Yimo Yan, Yong-Hong Kuo, and Zuo-Jun Max Shen
Accessed: 2023-6-23. 31 Qihao Wu, Jiangxue Han, Yimo Yan, Yong-Hong Kuo, and Zuo-Jun Max Shen. Reinforcement learning for healthcare operations management: methodological framework, recent developments, and future research directions.Health Care Management Science, 28(2):298–333,
2023
-
[2019]
Chan, Michael Huang, and Vahid Sarhangian
Carri W. Chan, Michael Huang, and Vahid Sarhangian. Dynamic server assignment in multiclass queues with shifts, with applications to nurse staffing in emergency departments.Operations Research, 69(6):1936–1959,
1936
-
[2021]
Chan, and Jing Dong
Yue Hu, Carri W. Chan, and Jing Dong. Prediction-driven surge planning with application to emergency department nurse staffing.Management Science, 71(3):2079–2126,
2079
-
[2023]
doi: 10.1287/opre.2022.2405. Benjamin Linthicum. Improving Emergency Department Throughput by Adoption of an Admissions Predictor Tool at Triage,
arXiv 2022
-
[2026]
doi: 10.1287/msom.2023.0700. Jim G. Dai and Pengyi Shi. Inpatient overflow: An approximate dynamic programming approach. Manufacturing & Service Operations Management, 21(4):894–911,
arXiv 2023
This paper was first reviewed by deepseek-v4-flash on August 1, 2026.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.