Operational Delay Agent

A synthetic hospital-flow app for finding out-of-bounds discharge delays, grouping them into plain-language delay reasons, and turning them into dashboard-ready action lists.

LIVE DEMO: Open the interactive dashboard

Demo source · Metric definitions

Operational Delay Sentinel is a small, self-contained prototype for hospital operations teams who need to answer a deceptively hard question:

Which discharge delays are no longer normal clinical variation, and which delay reason should we look at first?

It generates synthetic hospital-flow data, flags out-of-bounds operational delay signals, attributes likely operational delay reasons, trains local statistical baselines, and produces an interactive HTML dashboard plus CSV worklists. It is intentionally dashboard-first and CSV-first: no EHR write-back, no PHI, no punitive language, and no dependency on a live hospital system.

The important design choice is signal compression: thousands of patient-level rows are rolled into 3-5 management signal groups per shift, while case counts and recoverable bed-hours are reported separately. A delay reason can be high-impact without being high-volume, and common service throughput is not treated as a crisis just because it is frequent.

Why this exists

Hospital discharge delays are not just statistical outliers. A five-day miss on predicted discharge timing may be caused by an ALC/community placement wait, Friday service closure, delayed imaging, late blood results, ECG access, transport coordination, or home-care confirmation. Treating all of that as model error hides the operational story.

This app reframes those misses as reviewable delay signals:

• soft operational language,
• ranked recoverable excess bed-hours,
• clear recommended owners,
• auditable CSV action lists,
• dashboard filters for facility, unit, signal family, and exact delay reason.

Screenshots

Balanced one-hospital run, mixed signals with weekend/community pressure

Diagnostics-heavy run, detected as diagnostics-access dominant

Weekend-flow run, detected as Friday/weekend gap dominant

GitHub Pages demo

A cleaner public demo lives in demo/index.html. It is designed for GitHub Pages and tells one simple story: how a large synthetic hospital turns noisy discharge-delay evidence into a small, plain-language action agenda.

The demo uses committed JSON extracts generated from sample_data/, so it does not need Python, Parquet, a backend service, or real patient data in the browser. To refresh it after updating sample data, run:

python3 scripts/build_demo.py

To publish it with GitHub Pages, set Pages to serve from the main branch and open /demo/. The static page includes scenario tabs, executive KPI cards, delay-reason impact charts, a control-chart view, and example CSV-ready action rows.

What it detects

The dashboard presents operational diagnostic KPIs. These are not clinical quality judgements or staff performance ratings; they are review prompts that help patient-flow teams decide what to look at first.

Core KPI definitions

KPI	Formula	What it means	How to use it
Patient encounters	count(patient_admission_events.encounter_id)	Number of synthetic admissions/encounters in the run.	Denominator for signal rates.
Raw OOB delay signals	count(out_of_bounds_delay_flags where oob_flag = true)	Encounter-level rows that crossed at least one out-of-bounds rule.	Audit evidence, not the huddle worklist.
OOB signal rate	raw_oob_delay_signals / patient_encounters	Share of encounters with an out-of-bounds delay signal.	Overall operating pressure indicator.
Management signal groups	count(management_signal_groups)	Shift-level grouped signals after compressing raw patient rows.	Main huddle/action surface.
Average groups per shift	management_signal_groups / total_shift_windows	Typical number of signals a shift team reviews.	Keep manageable, usually under 3-5.
Delay-reason case rate	flagged_cases_for_delay_reason / patient_encounters	How often a specific reason appears.	Prevents overreacting to normal high-volume work.
Recoverable excess bed-hours	sum(estimated_recoverable_bed_hours)	Estimated bed-hours above expected operational thresholds.	Measures impact, not case count.
Bed-hour share by signal family	family_recoverable_bed_hours / total_recoverable_bed_hours	Concentration of impact in one family.	Used to label scenarios as mixed or dominant.
Compression ratio	raw_oob_delay_signals / management_signal_groups	How much raw evidence is compressed into huddle signals.	Shows whether the worklist is manageable.
Priority score	0.65 * hours_above_limit + 0.55 * post_ready_excess_hours + 18 * post_ready_hard_cap_flag + 8 * control_chart_signal_flag	Sorting score for operational review.	Higher means review sooner.

Core flag formulas

robust_los_oob_flag = actual_los_hours > oob_limit_hours
post_ready_hard_cap_flag = hours_after_medically_ready > 48
control_chart_signal_flag = daily_oob_rate > upper_control_limit
oob_flag = robust_los_oob_flag OR post_ready_hard_cap_flag OR control_chart_signal_flag

post_ready_excess_hours = max(hours_after_medically_ready - 48, 0)
los_excess = max(actual_los_hours - oob_limit_hours, 0)

if post_ready_excess_hours > 0:
    estimated_recoverable_bed_hours = min(los_excess, post_ready_excess_hours)
else:
    estimated_recoverable_bed_hours = 0.35 * los_excess

Delay reason KPI dictionary

Delay reason / KPI	Signal family	Definition	Formula / detection evidence	Default owner	Caveat
ALC placement wait	alc_community_capacity	Medically stable patient appears to be waiting for LTC, rehab, or alternate-level placement.	ALC status plus LTC/rehab referral or placement timing; KPI rate = ALC placement wait cases / encounters.	Transition services lead	High-impact but should be a small subset, not a dominant volume.
Friday/weekend discharge gap	weekend_flow_gap	Discharge readiness crosses a period with reduced weekend or Friday-afternoon service availability.	Medically ready near Friday/weekend plus delayed discharge; KPI rate = Friday/weekend cases / encounters.	Site operations director	Not every weekend discharge is a problem; public demo caps this near 5% or less.
Home-care confirmation wait	alc_community_capacity	Discharge depends on home-care confirmation or community support setup.	Home-care referral and confirmation timestamps; KPI rate = home-care wait cases / encounters.	Home-care liaison	Can overlap with frailty, ALC, and social-support needs.
Therapy assessment stall	care_team_assessment	PT/OT or therapy assessment takes longer than expected for a discharge-dependent case.	Assessment requested/completed timestamps above expected window.	Therapy services manager	Some delays reflect clinically appropriate assessment complexity.
Transport delay	transport_discharge_logistics	Discharge or transfer waits on transport coordination.	Transport requested/completed timestamps or discharge-order-to-transport lag.	Patient transport coordinator	May reflect external ambulance, family pickup, or destination constraints.
Pharmacy discharge delay	pharmacy_discharge_readiness	Medication reconciliation or discharge medication readiness delays discharge.	Discharge order timing plus pharmacy dependency timing.	Pharmacy operations lead	Complex medication reviews may be appropriate and safety-protective.
Radiology CT turnaround delay	diagnostics_access	CT completion or reporting appears discharge-dependent and delayed.	CT ordered/completed timestamps above expected discharge-dependent window.	Radiology operations lead	Frequent CT use is normal; only threshold-crossing cases signal.
Radiology MRI access delay	diagnostics_access	MRI access, completion, or reporting appears discharge-dependent and delayed.	MRI ordered/completed timestamps above expected window.	Radiology operations lead	MRI scarcity may be structural, not locally fixable same day.
Radiology ultrasound turnaround delay	diagnostics_access	Ultrasound completion or reporting appears discharge-dependent and delayed.	Ultrasound ordered/completed timestamps above expected window.	Radiology operations lead	Prioritization may be clinically appropriate.
Blood testing turnaround delay	diagnostics_access	Blood collection, processing, or result release appears discharge-dependent and delayed.	Blood test ordered/available timestamps above expected window.	Laboratory operations lead	Repeat or abnormal tests may be clinically necessary.
ECG availability delay	diagnostics_access	ECG completion or interpretation appears discharge-dependent and delayed.	ECG ordered/completed timestamps above expected window.	Cardiology diagnostics lead	Clinical urgency and interpretation requirements matter.
Diagnostic sign-off stall	diagnostics_access	Final diagnostic completion, sign-off, or consultant interpretation appears to hold discharge.	Diagnostic completion/sign-off timing above expected window.	Diagnostics operations lead	May reflect specialist dependency rather than diagnostic department delay.
Unit-level bed-flow bottleneck	unit_flow_capacity	A facility/unit/day has an unusual concentration of OOB delay signals.	daily_oob_rate > centerline + 3 * sigma.	Unit operations manager	Pattern signal only; not an individual-case conclusion.
Triage screening backlog	front_door_screening	Front-door screening delay may contribute to downstream flow pressure.	Triage screening requested/completed timestamps above expected window.	Triage operations lead	Flow signal, not discharge-only signal.
Nurse discharge screening wait	nursing_discharge_readiness	Nursing discharge readiness checklist or screening is delayed for a discharge-dependent case.	Nurse screening requested/completed timestamps above expected window.	Nursing flow lead	Staffing and acuity context should be reviewed before action.
Vulnerable patient porter wait	vulnerable_patient_flow	Mobility-limited or vulnerable patients wait for safe movement or porter support.	Porter request/completion timestamps plus frailty/mobility context.	Patient transport coordinator	Should be handled as access and safe-flow support, not waste.
Social work assessment wait	social_support_readiness	Social work assessment is delayed for a discharge-dependent case.	Social work assessment requested/completed timestamps above expected window.	Social work lead	May indicate complexity rather than avoidable delay.
Interpreter availability wait	social_support_readiness	Interpreter support is delayed for discharge-dependent communication.	Interpreter requested/completed timestamps above expected window.	Language services lead	Equity and access support should not be sacrificed for speed.
Discharge documentation readiness wait	documentation_readiness	Required discharge documentation is not ready in expected window.	Documentation started/ready timestamps above expected window.	Unit clerk lead	Documentation quality should not be reduced to improve speed.
Unattributed operational delay	unattributed	OOB delay exists but no single dependency explains it.	OOB flag without clear event-level attribution.	Patient flow coordinator	Often points to missing workflow instrumentation or data quality gaps.

How the sentinel works

flowchart LR
  A["Synthetic admissions"] --> B["Journey events"]
  B --> C["OOB delay detection"]
  C --> D["Delay reason attribution"]
  D --> E["Ranked signals"]
  E --> F["Shift-level management signal groups"]
  E --> G["Dashboard + CSV exports"]
  C --> H["Model baselines"]

Loading

The detection layer uses three complementary signals:

• robust length-of-stay control limits by facility, service line, case mix group, and frailty band,
• post-medically-ready hard caps,
• daily control-chart signals for systemic unit/facility patterns.

The action layer keeps three views:

• delay_blocker_attribution.parquet: raw patient-level delay-reason evidence,
• management_signal_groups.csv: capped shift-level management agenda, default 5 groups per shift,
• delay_resolution_actions.csv: supporting patient examples for the selected management groups.

Quick start

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Run the one-large-hospital balanced synthetic case:

python3 run_discharge_delay_workflow.py \
  --facilities 1 \
  --days 90 \
  --encounters-per-day 110 \
  --oob-rate-target 0.05 \
  --post-ready-hard-cap-hours 48 \
  --weekend-service-reduction 0.35 \
  --alc-pressure-multiplier 1.25 \
  --scenario-mode balanced \
  --out outputs/synthetic_90d_large_hospital_balanced_v1 \
  --print-top-n 5

Run the diagnostics-heavy and weekend-flow scenarios:

python3 run_discharge_delay_workflow.py --scenario-mode diagnostics_heavy --out outputs/synthetic_90d_large_hospital_diagnostics_v1
python3 run_discharge_delay_workflow.py --scenario-mode weekend_flow_gap --out outputs/synthetic_90d_large_hospital_weekend_v1

Scenario modes

Mode	Purpose
balanced	Mixed operational delay pressure. Useful default demo.
alc_heavy	ALC, LTC, rehab, home care, and community-capacity pressures dominate.
diagnostics_heavy	Radiology, blood testing, ECG, and diagnostic sign-off delays dominate.
weekend_flow_gap	Friday/weekend service gaps dominate recoverable bed-hours.

The app also detects the scenario actually produced by the data in scenario_detection_summary.csv. If no single signal family contributes at least 40% of recoverable excess bed-hours, the run is labelled mixed instead of forcing a misleading dominant category.

Latest synthetic run results

All three latest runs use one large synthetic hospital, 90 days, and 9,900 admissions. Earlier health-authority-scale stress runs are intentionally not used as the public demo baseline.

Scenario mode	Detected scenario	Raw OOB signals	Management groups	Avg groups / shift	Compression	ALC cases	ALC case rate	Weekend cases	Weekend case rate	Top signal family	Family share
balanced	mixed_operational_pressure_detected	811	543	2.25	1.49x	43	0.43%	119	1.20%	weekend_flow_gap	34.19%
diagnostics_heavy	diagnostics_heavy_detected	1,467	745	2.79	1.97x	7	0.07%	55	0.56%	diagnostics_access	54.95%
weekend_flow_gap	weekend_flow_gap_detected	910	456	1.85	2.00x	15	0.15%	466	4.71%	weekend_flow_gap	91.42%

The latest runs intentionally separate raw evidence volume, case rates, and recoverable excess bed-hours. Raw OOB rows remain available for audit, but the dashboard caps the shift-level agenda to a manageable number of signal groups. This prevents normal large-system friction, such as radiology throughput at scale, from being mislabeled as a crisis simply because it is frequent. ALC/community-capacity rows are bounded as a small high-impact subset rather than a structurally dominant source of synthetic delay.

Full run summary:

• docs/scenario_run_summary.csv
• docs/model_metrics_summary.csv
• docs/METRICS_AND_SIGNALS.md

Signal-volume guardrails

The public demo is tuned so signal volumes stay believable for one large hospital:

• no delay-reason group should represent more than about 5% of all encounters, even in a stress scenario,
• most delay reasons should sit around 1% to 5% of encounters or lower,
• common throughput processes, such as radiology or blood testing, are treated as delay reasons only when they cross timing and actionability thresholds,
• management views stay grouped to a small number of shift-level signals rather than exposing every patient-level row as an executive action.

This keeps the dashboard useful for a huddle: it points to a few practical signals, not a wall of normal operating friction.

One-hospital interpretation

The public demo is intentionally sized like a single large hospital, not a provincial or health-authority-wide extract. The current default produces:

• 9,900 admissions over 90 days,
• 811 to 1,467 raw OOB delay signals depending on scenario,
• 456 to 745 shift-level management groups,
• about 1.85 to 2.79 management groups per shift,
• ALC case rates between 0.07% and 0.43%.

That distinction matters: ALC is represented as a small, high-impact subset. Recoverable excess bed-hours are not case counts.

Metrics and signal definitions

The dashboard metrics are operational diagnostic KPIs. They are documented in detail in docs/METRICS_AND_SIGNALS.md, including formulas for OOB flags, priority score, recoverable excess bed-hours, management signal grouping, and every delay reason shown in the dashboard.

Dashboard features

The generated dashboard includes:

• KPI cards,
• recoverable excess bed-hours bar chart,
• OOB trend SVG chart,
• scenario detection mix,
• ranked actionable signals,
• executive worklist,
• control-chart daily metrics,
• filters for facility, unit, signal family, and exact operational signal,
• clickable bars that filter the tables,
• visible-row CSV export buttons,
• print/save-PDF support.

Dashboard screenshot/export

Install screenshot support:

pip install '.[screenshot]'
python3 -m playwright install chromium

Export dashboard HTML and PNG:

python3 scripts/export_dashboard.py \
  --dashboard outputs/synthetic_90d_large_hospital_weekend_v1/operational_delay_dashboard.html \
  --out exports \
  --png

GitHub repository structure

.
├── README.md
├── pyproject.toml
├── requirements.txt
├── run_discharge_delay_workflow.py
├── scripts/
│   └── export_dashboard.py
├── src/discharge_delays/
│   └── workflow.py
├── docs/
│   ├── scenario_run_summary.csv
│   ├── model_metrics_summary.csv
│   └── screenshots/
│       ├── balanced-dashboard.png
│       ├── diagnostics-dashboard.png
│       └── weekend-dashboard.png
├── sample_data/
│   ├── balanced/
│   ├── diagnostics_heavy/
│   └── weekend_flow_gap/
└── .github/workflows/ci.yml

Generated outputs/, exports/, local data, virtual environments, timestamped screenshot exports, and local implementation notes are ignored. The repository keeps only curated screenshots and summary CSVs for the README.

Included sample synthetic data

The repository includes a curated sample dataset under sample_data/. It contains the three final one-large-hospital scenario runs used in the README:

Folder	Scenario	Scale
sample_data/balanced/	Mixed operational pressure	1 hospital, 90 days, 9,900 encounters
sample_data/diagnostics_heavy/	Diagnostics-access stress	1 hospital, 90 days, 9,900 encounters
sample_data/weekend_flow_gap/	Weekend-flow stress	1 hospital, 90 days, 9,900 encounters

The larger local outputs/ tree is intentionally ignored because it contains earlier exploratory and health-authority-scale runs.

Main outputs

File	Purpose
patient_admission_events.parquet	Synthetic admission-level table.
patient_journey_events.parquet	Synthetic event-level journey table.
bed_resource_daily.parquet	Bed occupancy and resource context.
service_availability.parquet	PT, OT, imaging, pharmacy, home care, transport, LTC, rehab availability.
out_of_bounds_delay_flags.parquet	OOB delay signals and priority scores.
delay_blocker_attribution.parquet	Likely delay-reason attribution and evidence.
ranked_actionable_signals.csv	Raw ranked facility/unit/service/signal pattern table.
management_signal_groups.csv	Capped 3-5-per-shift management agenda.
management_signal_kpis.csv	Signal compression and manageability KPIs.
delay_resolution_actions.csv	Supporting patient-level examples for management groups.
delay_resolution_actions_all.csv	Full raw action inventory.
scenario_detection_summary.csv	Detected scenario mix by signal family.
operational_delay_dashboard.html	Interactive local dashboard.
discharge_delay_sentinel_report.md	Markdown run report.
discharge_delay_sentinel_report.html	HTML run report.

Model baselines

The workflow trains local statistical baselines:

• HistGradientBoostingRegressor,
• ExtraTreesRegressor,
• HistGradientBoostingClassifier,
• ExtraTreesClassifier.

The models are not the whole product. They are used to estimate expected LOS and OOB risk, while the operational layer converts delay signals into explainable delay reasons and worklists.

Language and governance

This project deliberately avoids punitive terminology. It uses terms like:

• delay signal,
• delay reason,
• capacity constraint,
• unresolved discharge dependency,
• recoverable excess bed-hours.

The intended first deployment pattern is shadow mode:

1. generate dashboard and CSV outputs,
2. review with discharge huddles and patient-flow teams,
3. record adoption and reasons-not-actioned,
4. only later consider idempotent integration with operational systems.

Synthetic data only

This repository uses fully synthetic data. It contains no PHI and makes no claim about a specific real hospital, health authority, or provincial program.

Cute but serious roadmap

• Add a proper web front end around the generated dashboard.
• Add adoption simulation: accepted, deferred, already resolved, not actionable.
• Add idempotent recommendation IDs for repeated daily runs.
• Add scenario comparison pages.
• Add optional integration adapters for CSV/SFTP/data-warehouse handoff.

License

MIT.