Table 1.
Previous studies predicting discharge within a fixed time window
| Reference | Population | Outcome | Model | Performance, AUROC where available |
|---|---|---|---|---|
| 7 | Adult patients discharged from inpatient surgical care in the US from May 1, 2016, to August 31, 2017; 15,201 hospital discharges | Discharge within 24 h | Multilayer perceptron neural network | AUROC 0.84 |
| 8 | Adult surgical patients discharged from inpatient care between July 2018 and February 2020; 10,904 patients during 12,493 inpatient visits | Discharge within 48 h | RF | AUROC 0.81 |
| 9 | Inpatients with cardiovascular diseases admitted to Asan Medical Centre in Korea between 2000 to 2016; 669,667 records | Discharge within the next 72 h (predictions not made on the day of discharge) | Extreme gradient boosting (XGB) | AUROC 0.87 |
| 10 | Adult patients admitted to Vanderbilt University Medical Centre in 2019; 26,283 patients | Discharge within 24 h | Light gradient boosting machine (LGBM) |
AUROC 0.92 with user-EHR interactions; AUROC 0.86 without user-EHR interactions. |
| 11 | Patients admitted to a mid-Atlantic academic medical centre from 2011-2013; 8852 patient visits and 20,243 individual patient days | Discharge within 7 and 17 h (from 7 am) |
Logistic regression (LR); Random Forest (RF) |
Sensitivity: LR: 65.9; RF: 60.0; Specificity: LR: 52.8; RF: 66.0 |
| 12 | Adult patients admitted to a community hospital in Maryland, USA between April 2016 and August 2019; 120,780 discharges for 12,470 patients | Discharge on the same day, by the next day, within the next 2 days | RF |
AUROC 0.80 (same day); AUROC 0.70 (next day) |
| 13 | Inpatients admitted at Beth Israel Deaconess Medical Centre between January 2017 and August 2018; 63,432 unique admissions (41,726 unique patients) | Discharge within 1 day, discharge within 2 days | LR, CART decision trees, Optimal trees, RF, Gradient boosted trees |
Discharge within 1 day, AUROC 0.84; Discharge within 2 days, AUROC 0.82 |
| 14 | Patient encounters from 14 different Kaiser Permanente facilities in northern California from November 1, 2015 to December 31, 2017; 910,366 patient-days across 243,696 patients hospitalisations | Discharge within 1 day | LR, Lasso, RF, GBM | GBM, AUROC 0.73 |
We searched Google Scholar and PubMed for studies up to 30 April 2024, using the search terms ‘machine learning’ AND (‘hospital discharge prediction’, OR ‘patient flow’). AUROC: area under the receiver operating curve. XGB: Extreme gradient boosting. GBM: gradient boosting machine. LR: logistic regression. RF: random forest. CART: classification and regression tree. The features used in each model are summarised in Supplementary Table 1.