Development of AI Hospital Load Forecasting System
Patient flow forecasting in hospital — task with direct impact on care quality. Staff shortages during peaks lead to treatment delays and errors, excess causes inefficient spending. Accurate load model enables shift planning, operating rooms and supply management.
What to Forecast
Incoming Flow:
- Emergency / admission department visits: by hours, days
- Planned hospitalizations: by specialties (surgery, therapy, cardiology)
- Discharges: for bed release
- Length of Stay (LOS): for available bed calculation
Resource Requirements:
- Nursing staff needs by departments
- Operating room utilization
- Consumables need (ventilators, medications)
- Laboratory testing needs
Factors Affecting Load
| Factor | Examples | Influence Horizon |
|---|---|---|
| Seasonality | Flu in winter, injuries in summer | Weeks/months |
| Day of Week | Monday peak visits, Sunday minimum | 1-7 days |
| Holidays | New Year — injury surge | Specific dates |
| Weather | Frost → cardiovascular | 1-3 days |
| Epidemiology | ARVI waves, outbreaks | 1-4 weeks |
| Demographics | Aging population | Years |
Weather data (temperature, humidity, atmospheric pressure) — statistically significant predictors for cardio and pulmonology departments.
Forecasting Models
Basic Approaches:
- SARIMA: good for regular patterns (weekly + annual seasonality)
- Prophet: simplicity and interpretability, holiday regressors
ML Approaches:
- LightGBM with lag features + weather + epidemiological indices (Flu Index from Yandex/Google)
- Gradient Boosting shows MAPE 6-10% for weekly horizon
Specifics:
- Can't allow look-ahead bias in training
- Epidemiological data delay: flu index published with 1-2 week lag, need proxy features (search queries)
- Missing data: zero values at night ≠ missing data
Length of Stay (LOS) Forecasting
LOS prediction enables bed availability calculation:
LOS Model Features:
- Diagnosis (ICD-10 code)
- Age, gender, comorbidities (CCI — Charlson Comorbidity Index)
- Admission type (planned / emergency)
- Department
- Initial exam data (lab values at admission)
Algorithm: Survival analysis (AFT — Accelerated Failure Time) or quantile regression. Metric: MAE in days (typically 1.5-2.5 days with average LOS 5-7 days).
Resource Planning
Staffing Calculation:
Nurses_needed(hour) = ceil(Expected_Patients(hour) / Nurse_Patient_Ratio)
Nurse_Patient_Ratio: depends on department (ICU: 1:2, therapy: 1:8)
Operating Rooms: Forecast planned operations by specialty + emergency urgency (history + seasonality). OR schedule optimization via CP-SAT considering teams, equipment, procedure duration.
Supplies: Consumables forecast = f(expected patient volume, procedure mix). Integration with pharmacy system (PHARMACY 2.0, 1C:Hospital Pharmacy) for automatic requests at ROP.
Dashboard and Alerts
Operational Dashboard (for chief physicians and senior nurses):
- Current load vs. forecast by departments in real-time
- Forecast for next 24/48/72 hours
- Alert on expected staff/bed shortages
Strategic Dashboard (for administration):
- Forecast accuracy metrics
- Seasonal patterns year over year
- Capacity planning: need additional hospital capacity
MIS Integration: EMIAS (Moscow), SAMSON, MedElement, 1C:Hospital — API or HL7 FHIR for admission/discharge data.
Timeline: basic admission forecast + shift planning — 8-10 weeks. Full system with LOS forecast, OR scheduling and dashboard — 4-5 months.