Ishanu Chattopadhyay PRO
ML | Data Science Biomedical Informatics | Social Science | Assistant Professor
David Yang
Ishanu Chattopadhyay
05072026
Dandelion Dataset
~15117 patients ~6 months
- synchronization issues (discharge before admission)
- time stamps often too coarse
- No procedure durations
- dx codes, procedural codes, admission and discharge times,
- course ADT data,
- patients often have multiple encounters recorded, noisy service department information
Predict Length of Stay*
*Use only information available at/before admission
Models: LGBMs
Length of Stay
Length of Stay
Take-away:
if a patient is going to stay longer than a day is highly predictable, and predictability falls with longer cut-points
Length of Stay: Aggregate to Model Patient Load
R^2 > 90\%
Length of Stay: Aggregate to Model Patient Load
Take-away: Very predictable
Predict Procedures*
*Use only information available at/before admission
Take-away: Not as predictable, but not bad
Take-away: Aggregates are more predictable
Predict Procedure Load (OOS)
Take-away: Very predictable as a time series
| Procedure | Expected time (min) | Estimated SD (min) | Notes |
|---|---|---|---|
| SUBSEQ HOSP CARE MOD MDM 35 MIN | 35 | 10 | Time-based inpatient follow-up; moderate variation |
| ELECTROCARDIOGRAM ROUTINE WAT LEAST 12 LEADS INTERP & RPT ONLY | 7 | 3 | Short interpretation task; usually low spread |
| EMERGENCY DEPT VISIT LVL V | 75 | 30 | High-complexity ED visit; broad variability |
| NOTE RECON VERIFICATION | 10 | 5 | Documentation/reconciliation task; operational estimate |
| HOSP DSCHG DAY MGMT >30 MIN DATE OF ENCTR | 40 | 12 | Discharge management above threshold; moderate spread |
| INIT HOSP CARE MOD MDM 55 MIN | 55 | 15 | Initial inpatient evaluation; broader than follow-up |
| INIT HOSP CARE HIGH MDM 75 MIN | 75 | 20 | Complex initial inpatient care; moderate-high spread |
| SUBSEQ HOSP CARE HIGH MDM 50 MIN | 50 | 15 | High-complexity follow-up; moderate spread |
| HOSP DSCHG DAY MGMT <30 MIN DATE OF ENCTR | 20 | 8 | Shorter discharge workflow; moderate spread |
| EMERGENCY DEPT VISIT LVL IV | 45 | 18 | Moderate-high complexity ED visit; fairly variable |
Approximate Procedure Times (LLM)
Predict Procedure Daily Load in minutes
Take-away: Very predictable as a time series
- Length of Stay
- Patient Load over time
- Procedural load over time
What else can we do with this data:
- Use predictive models of future risk for individual patients (using models learned elsewhere)
- Make use of detailed ICD code data to inform models
- Model patient journey network with LLM-based ADT reverse-engineering
Dandelion Analysis
By Ishanu Chattopadhyay
Dandelion Analysis
Dandelion Data Modeling
