Predictive algorithms are increasingly used to guide hospital discharge decisions, yet these models can exhibit censoring bias for vulnerable patients due to insufficient historical data or past discharge decisions. When biased readmission risk predictors are used to inform discharges, the training data generated by discharge decisions may perpetuate similar biases when predictors are re-trained, leading to a vicious cycle. We extend a queueing-based Markov decision process for hospital discharge to incorporate periodically retrained readmission risk predictors. Our key insight is that cost-optimal, "capacity-aware" operational decision policies can naturally diversify the lengths of stay at which patient outcomes are observed, enabling consistent estimation of readmission risk without explicit fairness constraints. We prove that repeated retraining under discharge-up-to policies converges to unbiased risk estimates under well-specification. In synthetic experiments and calibrated simulations based on patient-level data from a partner hospital, we find that where commonly used threshold-on-risk policies can remain persistently biased, discharge-up-to policies reduce learning error over status quo decision policies (such as thresholding on risk) by 24%-25%. These improvements to estimation also mitigate some forms of bias, reducing readmission cost disparities (adjusted for baseline risk differences between demographic groups) by 38%-63%. Our results demonstrate how capacity-aware decision policies can improve both predictive performance and fairness by addressing censoring bias.
Figures & tables
Figure 1: (a) Distribution of discharge decisions made by the discharge-up-to policy and baselines over 100,000 days. Policies decide discharges using a univariate logistic model on LOS with r^(1)=0.7 , r^(8)=0.3 . (b) rMSE of the best and worst-performing parameter in each policy class at the final episode for a representative risk model configuration ( r(1)=r^1(1)=0.7 , r(8)=r^1(8)=0.3 ). (c) Number of censored days in the final episode vs. the simple threshold policy’s rMSE . (d) Readmission cost vs. state aware policy’s rMSE and number of censored days in the final episode.
Figure 2: Performance of the discharge up to policy compared to best-performing baselines in a simulation calibrated with real hospital readmission data. (a) rMSE of all model iterates with bootstrapped 95% confidence intervals. (b) Total cost of each policy broken down into readmission and congestion cost, for each cost regime. SE bars are for the total cost. (c) rMSE by group (Caucasian and African American) at the final episode, for the best-performing parameter of each policy class. (d) Readmission disparities and baseline-readjusted disparities (Caucasian - African American) at the final episode in each cost regime.
Method
rMSE
∣ΔrMSE∣
Cost
∣ΔcB∣
Dis* ( D~1=30,cR=5 )
0.095
0.0113
558.9974
0.0538
Thresh
0.1212 ( τ=0.3 )
0.0110 ( τ=0.2 )
854.8178
0.0862 ( τ=0.1 )
Improvement over Thresh
24.4% ( p<.001 )
-2.49% ( p=.59 )
34.4% ( p<.005 )
40.3% ( p<.005 )
State ( cR=5 )
0.123
0.0114
1212.6527
0.1439
Improvement over State
25.5% ( p<.005 )
0.8% ( p=.48 )
37.7% ( p<.005 )
62.7% ( p<.005 )
Table 1: Comparison of performance of discharge-up-to policy at the final episode cR=5 against best-performing baselines.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Notation
Description
∥⋅∥1 , ∥⋅∥
Entrywise L1 norm
Ak,w
Number of arrivals on day k for patients of type w
Ak
Number of arrivals of all types, ∑w∈WAk,w
cB(Dk)
Expected readmission cost of discharges adjusted for baseline risk (risk if held until maximum LOS), cR∑w=1M∑ℓ=1JDk,wℓ⋅(rw(ℓ)−rw(J))
cB(Dk,w)
Expected readmission cost of discharges from group w adjusted for baseline risk cR∑ℓ=1JDk,wℓrw(ℓ)
ΔcB(i,w1,w2)
Disparity in adjusted expected readmission cost between w1 and w2 for round i . See Appendix C.1 for definition.
Appendix
Table 2: Notation table.
Figure 3: Readmission risk as a function of LOS for two groups. For group 1, the logit ln(r(ℓ)/(1−r(ℓ))) follows −1.6ℓ+3 ; for group 2, it follows −ℓ+0.8 . Group 1 has higher baseline risk and larger marginal risk reductions from continued hospitalization; group 2 has lower baseline risk and smaller marginal reductions.
Figure 4: (a) Average risk over the empirical distribution of patients for the realistic model per LOS. (b) Average risk for three k -means clusters of patient risk trajectories. (c) Risk for 50 randomly sampled patients from our dataset.
Figure 5: Average of biased model predictions vs. true risk predictions for the empirical distribution of patient covariates conditional on race.
Policy Name
Mean Total Cost ( h=1,cR=100 )
SEM Total Cost
Mean rMSE
SEM rMSE
Dis ( D~=10 )
5.97×104
180.3612
0.0117
0.0002
Dis ( D~=20 )
8.49×104
170.6799
0.0082
0.0002
Dis ( D~=30 )
1.21×105
177.2092
0.0067
0.0001
Dis ( D~=40 )
1.68×105
188.2904
0.0062
0.0001
Dis ( D~=50 )
2.24×105
209.0363
0.0072
0.0001
Dis ( D~=60 )
2.86×105
327.2783
0.0098
0.0002
Appendix
Table 3: Sample means and SEs for cost and rMSE of each policy at the final episode across 50 true/initial configurations of risk.
Policy Name
Mean Cost ( h=1,cR=100 )
SEM Total Cost
Mean rMSE
SE
Dis ( D~=10 )
5.98×104
588.5656
0.0118
0.0015
Dis ( D~=20 )
8.48×104
435.4185
0.0093
0.001
Dis ( D~=30 )
1.21×105
652.0527
0.007
0.0006
Dis ( D~=40 )
1.68×105
819.1144
0.0067
0.0009
Dis ( D~=50 )
2.24×105
1006.7915
0.0076
0.0008
Dis ( D~=60 )
2.86×105
2295.5058
0.0108
0.0014
Appendix
Table 4: Sample means and SEs for cost and rMSE of each policy at the final episode for example risk configuration r(1)=r^(1)=0.7 , r(J)=r^(J)=0.3 .
Policy Name
Mean rMSE
SEM rMSE
Mean ΔrMSE
SEM ΔrMSE
Dis ( D~=10 )
0.0982
0.0008
-0.0037
0.0016
Dis ( D~=20 )
0.0845
0.0009
-0.0094
0.0016
Dis ( D~=30 )
0.0769
0.001
-0.0059
0.0012
Dis* ( D~1=10 , cR=125 )
0.0845
0.0009
-0.0043
0.0014
Dis* ( D~1=10 , cR=5 )
0.0978
0.0006
-0.01
0.0016
Dis* ( D~1=20 , cR=125 )
0.0817
0.0009
-0.0069
0.0013
Appendix
Table 5: Sample means and SEs for rMSE and rMSE disparities (Caucasian-African American) of each policy at final episode in realistic experiments.
Policy Name
Mean Cost ( cR=5 )
SEM Cost ( cR=5 )
Mean Cost ( cR=125 )
SEM Cost ( cR=125 )
Dis ( D~=10 )
541.6881
6.0977
1.33×104
146.7497
Dis ( D~=20 )
832.7707
7.7814
1.18×104
131.5977
Dis ( D~=30 )
1344.1148
8.3699
1.11×104
138.142
Dis* ( D~1=10 , cR=125 )
NA
NA
1.15×104
165.1165
Dis* ( D~1=10 , cR=5 )
555.0492
6.1148
NA
NA
Dis* ( D~1=20 , cR=125 )
NA
NA
1.12×104
183.4523
Appendix
Table 6: Sample means and SEs for cR and cR disparities (Caucasian-African American) of each policy at final episode in realistic experiments.
Policy Name
Mean ΔcR ( cR=5 )
SEM ΔcR ( cR=5 )
Mean ΔcR ( cR=125 )
SEM ΔcR ( cR=125 )
Dis ( D~=10 )
-0.1976
0.0332
-4.9409
0.8305
Dis ( D~=20 )
-0.2584
0.0353
-6.4595
0.8824
Dis ( D~=30 )
-0.1809
0.0325
-4.5226
0.8119
Dis* ( D~1=10 , cR=125 )
NA
NA
-3.8986
0.852
Dis* ( D~1=10 , cR=5 )
-0.2309
0.0443
NA
NA
Dis* ( D~1=20 , cR=125 )
NA
NA
-4.4796
1.033
Appendix
Table 7: Sample means and SEs for cB and cB disparities (Caucasian-African American) of each policy at final episode in realistic experiments.
Policy Name
Mean ΔcB ( cR=5 )
SEM ΔcB ( cR=5 )
Mean ΔcB ( cR=125 )
SEM ΔcB ( cR=125 )
Dis ( D~=10 )
0.0642
0.0078
1.6053
0.196
Dis ( D~=20 )
0.1278
0.0093
3.1949
0.2328
Dis ( D~=30 )
0.1513
0.01
3.7824
0.2498
Dis* ( D~1=10 , cR=125 )
NA
NA
2.7694
0.3119
Dis* ( D~1=10 , cR=5 )
0.0409
0.0086
NA
NA
Dis* ( D~1=20 , cR=125 )
NA
NA
2.893
0.2753
Appendix
Table 8: Sample means and SEs for cB and cB disparities (Caucasian-African American) of each policy at final episode in realistic experiments.
Policy 1
Policy 2
Metric
%
Decrease/Increase
p -value
Dis* ( D~1=30 , cR=5 )
Thresh ( τ=0.3 )
rMSE
33%
Dec.
1.33×10−29
Dis* ( D~1=30 , cR=5 )
State ( cR=5 )
rMSE
34%
Dec.
1.22×10−29
Dis* ( D~1=30 , cR=5 )
Thresh ( τ=0.3 )
∣ΔrMSE∣
11%
Dec.
0.53
Dis* ( D~1=30 , cR=5 )
State ( cR=5 )
∣ΔrMSE∣
23%
Dec.
0.47
Dis* ( D~1=20 , cR=5 )
Thresh ( τ=0.3 )
Cost
35%
Dec.
5.37×10−23
Dis* ( D~1=20 , cR=5 )
State ( cR=5 )
Cost
54%
Dec.
2.37×10−35
Appendix
Table 9: Full significance test results for percent improvements.
Figure 6: ERM vs. IPW rMSE for the discharge-up-to level policy in the misspecified setting.