The document discusses predicting patient discharge from hospital wards with no real-time clinical data. It presents a random forest model using features like ward admissions/discharges, patient characteristics, and time series patterns. The random forest model improved discharge predictions over baseline ARIMA and naive models, with a 25% reduction in error versus naive forecasts. Seasonality like day-of-week was found to influence predictions, suggesting administrative factors impact discharges. The proposed commonly available data-driven approach could integrate into hospital systems to help manage bed occupancy.
Forecasting patient outflows without clinical data
1. Forecasting patient outflow from wards having
no real-time clinical data
Shivapratap Gopakumar
Truyen Tran, Wei Luo, Dinh Phung, Svetha Venkatesh
PPattern RRecognition aand DData AAnalytics
School of Information Technology
Deakin University, Australia
ICHI’16
Chicago
2. Introduction
Demand for Healthcare services increasing
2
“There is growing concern in various countries that the methods of providing
health care services are, if not already, approaching a level that will not be
sustained by the population.”
Mackay 2005; WHO report; European Commission report
Inpatient beds reduced by 2% since the last decade
Increased levels of bed occupancy = high throughput to contain
costs
Efficient bed management is key to avoid bed crisis
3. Predicting discharge from ward
Little attention for predicting
discharges from general wards
3
Daily discharge rate = indicator of
efficiency
Ward Manager
Recovery ward
Current demand
Past experience
Number of beds needed
Can we provide a good estimate for total
next-day discharges from the ward?
Significance: Relieve emergency
access block !
4. Challenges
4
No real-time clinical data.
Case-mix of patients in
ward.
Non-linear hospital
dynamics.
Variation in data
Discharge pattern for each week
Each colour represents a week
5. Related Work
5
Majority of studies on flow
in Emergency department.
Other studies target wards
with real-time clinical data.
To the best of our
knowledge, this is the first
study for open ward with
no real-time clinical data
7. Data: Patterns
7
Weekly discharge pattern Monthly discharge pattern
Daily discharges Time series decomposed to:
• Trend: long time change in mean level
• Seasonality: seasonal variations in the data
• Noise
8. Baseline Model: ARIMA
Autoregressive integrated moving average (ARIMA)
8
able to capture trends and seasonal variations and update
the changes over time.
Forecasted Discharge at time t
sum of recent discharges sum of recent forecast errors
9. Our contribution:
Feature engineering and random forest
9
Random Forest: creates an ensemble of decision trees
Tree 1
Tree n
Tree bagging + random feature selection
= good prediction with great control on overfitting
10. 10
Our contribution:
Feature engineering and random forest
We derived three groups of features from Ward data:
Ward level, Patient level, Time series
Ward-level features:
Admissions: in past 7 days
Discharges: in past 7 days
Occupancy: in the previous day
11. 11
Our contribution:
Feature engineering and random forest
Patient-level features:
Type of admission: 5 categories
Unit referred from : 49 categories
Patient class: 21 categories
Age: 8 categories
# Wards visited: 4 categories
Elapsed length of stay for each patient
12. 12
Our contribution:
Feature engineering and random forest
Time-series features:
Seasonality: Current day-of-week, month, time-series
decomposition
Trend: Polynomial regression
13. Experiment
• Baseline models: ARIMA, Naïve forecast (median
discharge)
• Compared with Random forest with our feature set
13
14. Experiment: Measuring performance
14
Mean Forecast Error:
Mean Absolute Error:
Root mean square error:
Symmetric Mean Absolute Percentage Error:
= True discharge at t = Forecasted discharge at t
15. Results
Random forest predictions:
25% improvement over Naive forecasting
17% improvement over ARIMA
Least error for each day-of-week
15
RMSE
16. Discussion
16
Seasonality:
time-series decomposition
Number of patients in
ward, the previous dayPatients with only 1
ward visited before
current.
Number of males in
ward# dishcharges on prev
14th
dayForecasted trend using
polynomial regression
“Public Standard”
Discharges21 days
before
Elapse patient length
of stay
18. Conclusion
1. Pronounced weekly patterns, as discussed in other studies
suggests discharges are heavily influenced by
administrative reasons and staffing
1. Forecast performance is not as good as emergency/acute
care studies.
But no clinical data available.
1. Proposed model built from commonly available data.
Can be easily integrated into existing systems.
18
20. References
• A. Kalache and A. Gatti, “Active ageing: a policy framework.” Advances in gerontology, vol. 11, pp. 7–18,
2002.
• M. Mackay and M. Lee, “Choice of models for the analysis and forecasting of hospital beds,” Health Care
Management Science, vol. 8, no. 3, pp. 221–230, 2005.
• M. Connolly, C. Deaton, M. Dodd, J. Grimshaw, T. Hulme, S. Everitt, and S. Tierney, “Discharge
preparation: Do healthcare professionals differ in their opinions?” Journal of interprofessional care, vol. 24,
no. 6, pp. 633–643, 2010.
• M. V. Shcherbakov, A. Brebels, N. L. Shcherbakova, A. P. Tyukov, T. A. Janovsky, and V. A. Kamaev, “A
survey of forecast error measures,” World Applied Sciences Journal, vol. 24, pp. 171–176, 2013.
• J. S. Peck, J. C. Benneyan, D. J. Nightingale, and S. A. Gaehde, “Predicting emergency department inpatient
admissions to improve same-day patient flow,” Academic Emergency Medicine, vol. 19, no. 9, pp. E1045–
E1054, 2012.
• S. Barnes, E. Hamrock, M. Toerper, S. Siddiqui, and S. Levin, “Real-time prediction of inpatient length of
stay for discharge prioritization” Journal of the American Medical Informatics Association, 2015.
• M. J. Kane, N. Price, M. Scotch, and P. Rabinowitz, “Comparison of arima and random forest time series
models for prediction of avian influenza h5n1 outbreaks,” BMC bioinformatics, vol. 15, p. 276, 2014.
• W. Luo, J. Cao, M. Gallagher, and J. Wiles, “Estimating the intensity of ward admission and its effect on
emergency department access block,” Statistics in medicine, vol. 32, no. 15, pp. 2681–2694, 2013.
20
Our research is motivated by the rising demand for healthcare services.
There has been an increased demand all over the world as is reflected by different studies.
Population growth, increase in life expectancy, work force issues
BUT number of beds are reducing. Could be due to better preventive measures, better services.
Most studies focus on emergency/acute-care patient flow. These centers are rich in data and resources because hospital performance is measured using these indices.
Daily discharge rate can be a potential real-time indicator of operational efficiency [Wong et al., 2010 ]
Significance
Data: Only admin data, no clinical information about patients. No info about medications, procedures, current diagnosis
Patients come from different places: Direct admissions, from emergency, from other wards
Hospital procedures are nonlinear: waiting for some service, tests etc.
ARIMA is widely used for time series forecasting
Popular due to : ease of formulation and interpretability
AR term: forecasted discharge is regressed over previous discharges
MA term: forecast error is linear combination of past discharge errors
Ensemble method:
Random decision forests correct for decision trees' habit of overfitting to their training set.
RF can rank importance of features in a natural way.
This “weekend effect” could be attributed to shortages in staffing, or reduced availability of services like sophisticated tests and procedures
Future Work:
1. Incorporate nonlinear statistics of hospitals: holidays, staff planning/leaves, availability of tests etc.