55 Healthcare Analytics: Patient Outcomes and Operations

55.1 Why Healthcare Analytics Matters

A retail dashboard misread costs a markdown; a healthcare dashboard misread costs a life.

Healthcare is the function with the highest stakes and the longest analytical lineage. Hospitals have measured length of stay, mortality, and readmission since the 19th century — Florence Nightingale’s coxcomb chart of Crimean War deaths is one of the founding works of data visualisation. Modern healthcare analytics extends that lineage with electronic health records, claims data, sensor streams, and population datasets, and uses BI tools to make the resulting picture legible to clinicians, administrators, regulators, and patients.

For a BI analyst, healthcare clusters into three jobs that mirror Donald M. Berwick et al. (2008)’s Triple Aim — improving the experience of care, improving the health of populations, and reducing the per-capita cost of care. Patient outcomes analytics answers are patients getting better, and are we doing harm we could prevent? — mortality, readmission, complication, infection, length-of-stay. Operations analytics answers are the wards, theatres, and clinics running well? — bed occupancy, OR utilisation, ED door-to-doctor time, staffing rosters. Population health analytics answers which segments of our community need what? — chronic-disease registries, screening coverage, preventive care, cost stratification. David W. Bates et al. (2014) argue that the highest-leverage healthcare analytics product is the high-risk patient registry — using analytics to identify the small fraction of patients who account for a large share of cost and adverse outcomes — and that this is fundamentally a visualisation problem as much as a modelling one.

The healthcare-dashboard contract

Three rules separate healthcare dashboards from every other kind:

Patient privacy is not optional. PHI lives behind row-level security, audit logs, and aggregation thresholds. A dashboard cell with fewer than 5 patients is suppressed by default.
Risk-adjustment is mandatory for outcomes. Raw mortality means nothing without case-mix adjustment; the sickest hospitals always look worst before adjustment.
Clinicians read at the bedside. The dashboard runs on a tablet or wall display in front of patients and staff. Slow loads and small fonts kill adoption.

55.2 Patient Outcomes Analytics

Outcomes analytics tracks whether patients get better, worse, or stay the same after contact with the health system. Four families of outcome dominate clinical reporting: mortality, readmission, complications and adverse events, and patient-reported outcomes (PROs).

Four Outcome Families and Their Visualisations

Outcome family	Question	Standard metric	Visualisation lens
Mortality	Are patients dying when they should not?	Risk-adjusted in-hospital and 30-day mortality.	Funnel plot against expected mortality with control limits.
Readmission	Did the patient stabilise, or come back?	30-day all-cause readmission, condition-specific readmission.	Bar chart by DRG with reference line; heatmap by ward.
Complications and harm	Did we cause something we should have prevented?	HACs, hospital-acquired infections, falls, pressure ulcers.	SPC chart with upper control limit; sortable list of harm events.
Patient-reported outcomes	Does the patient feel better?	PROMs scores pre/post intervention.	Paired-line chart of PROM trajectories per condition.

Risk-adjustment and the funnel plot

A raw mortality rate of 3.2 percent looks bad until you discover the hospital takes the most complex transfers in the region. Risk-adjustment computes an expected mortality for each patient based on age, comorbidities, severity at admission, and procedure, then aggregates expected and observed values per provider.

The standard visualisation is the funnel plot: x-axis is expected mortality count (a proxy for case volume and case mix); y-axis is observed mortality rate; control limits (95 percent and 99.8 percent) fan out from the centre. Each point is a hospital or clinician. Points above the upper control limit are outliers worth investigating; points within the funnel are consistent with the network average. The plot replaces ranking-table league tables with statistical signalling and is the standard view used by NHS, CMS, and most national outcome registries.

flowchart LR
  A[Patient cohort] --> B[Risk model<br/>age, comorbidities,<br/>severity]
  B --> C[Expected mortality<br/>per patient]
  C --> D[Aggregate per provider]
  D --> E[Funnel plot<br/>observed vs expected]
  E --> F{Outside<br/>99.8% limit?}
  F -->|Yes| G[Investigate<br/>case review]
  F -->|No| H[Consistent<br/>with network]
  style G fill:#FCE8E6,stroke:#D93025
  style H fill:#E6F4EA,stroke:#137333

Readmission and the cohort retention curve, inverted

The 30-day readmission rate is the headline number, but the more useful view is an inverted cohort curve — for each discharge cohort, plot the cumulative percentage of patients readmitted at days 1, 7, 14, 30, 60, 90. The shape reveals whether readmission is concentrated in the first week (often a discharge-instructions or medication-reconciliation problem) or the third week (often a primary-care follow-up gap). The fix differs by shape, and the chart is what tells the team where to focus.

Beware the SHMI black box

The Summary Hospital-level Mortality Indicator (SHMI) and similar composite measures are widely cited but easy to misread. SHMI is sensitive to coding depth (better-coded hospitals look better even if outcomes are unchanged) and to palliative-care exclusions. Always show SHMI alongside its inputs — observed vs expected mortality, coding depth, palliative-care share — so the audience can interpret the headline.

55.3 Operations Analytics: Beds, Theatres, and Clinics

Healthcare operations analytics is the BI face of capacity management — making sure the right patient is in the right bed, with the right staff, at the right time. Three views run on every operations dashboard.

Bed occupancy and patient flow

The bed-occupancy heatmap shows wards (rows) by hour of day (columns) with cell colour = percent occupancy. Above 90 percent occupancy is the danger zone — admissions back up in the ED, electives are cancelled, infection risk rises. The chart, run live on a wall display, lets the bed-management team see the next 4-hour pinch point.

A second view, the patient-flow Sankey, traces the journey from ED admission through ward → ICU → step-down → discharge or transfer. Bottlenecks show up as narrow links where wide ones should be — the SDU to step-down handoff is a chronic culprit. The chart was popularised in NHS flow improvement work and remains the cleanest way to surface system-level constraints.

Operating theatre utilisation

OR utilisation analytics combines four metrics:

Block utilisation — actual surgery time in a surgeon’s allocated block over the block duration.
First-case on-time start (FCOTS) — percent of first cases that start within 5 minutes of scheduled.
Turnover time — minutes between one case ending and the next beginning.
Cancellation rate — percent of scheduled cases cancelled, with reason codes.

The dashboard view is a four-panel small-multiples per OR suite, plus a sortable surgeon-block-utilisation table. Sub-60-percent block utilisation is a scheduling-policy problem; over-90-percent is a safety problem. The dashboard is the input to the quarterly OR governance committee — without it, blocks default to historical entitlement rather than evidence.

Emergency Department dashboards and the Door-to-Doctor clock

ED dashboards live and die by time-to metrics. Door-to-triage, door-to-doctor, door-to-decision, and disposition-to-departure together describe the patient’s hour-by-hour experience. The standard visual is a small-multiples histogram of each interval’s distribution, with the 4-hour ED total and 30-day median plotted alongside. ED leaders use the chart to triage which time interval is the binding constraint at each shift change.

Run charts beat snapshots in operations

Healthcare operations data is noisy — staffing patterns, day-of-week, weather, public holidays. Always plot operational metrics as run charts (or SPC charts with control limits), not as single bars or cards. A 92 percent OR utilisation card looks healthy until the run chart reveals it has been steadily declining for 14 weeks. The trend is the story; the snapshot is decoration.

55.4 Population Health Analytics

Population health analytics moves from the individual patient to the registered population — usually a primary-care list, an insured panel, or a geographic catchment. The data is the same kind (claims, EHR extracts, registry feeds), but the unit of analysis is different: a cohort of people, watched over time.

The chronic-disease registry

A chronic-disease registry is a maintained list of patients with a defined condition (diabetes, hypertension, COPD, CKD) plus the metrics that matter for that condition (HbA1c, BP, FEV1, eGFR). The dashboard shows three things together: count of registered patients, percent meeting target on their key metric, and percent up-to-date on guideline-recommended care (annual eye exam for diabetes, statin if indicated, etc.). Sorted lists of care gaps — patients overdue for screening — drive the practice manager’s daily work.

Cost stratification and the high-risk pyramid

David W. Bates et al. (2014) popularised the high-risk pyramid: 5 percent of patients account for 50 percent of cost, and identifying that 5 percent before they incur the cost is a major analytical opportunity. The visualisation is a horizontal Pareto bar of patients sorted by predicted-12-month-cost, with cumulative-share line overlaid. Patients in the top 5 percent feed into care-management programmes; patients in the next 15 percent (the rising risk segment) feed into preventive interventions.

flowchart TB
  A[Top 5%<br/>High-cost / high-risk<br/>Care management] --> B[Next 15%<br/>Rising risk<br/>Preventive intervention]
  B --> C[Bottom 80%<br/>Population health<br/>Wellness, screening]
  style A fill:#FCE8E6,stroke:#D93025
  style B fill:#FFF7E6,stroke:#F4B400
  style C fill:#E6F4EA,stroke:#137333

Health equity and the disparities lens

The same outcome metric (mortality, readmission, screening coverage) cut by demographic dimension (income decile, ethnicity, geography, language preference) often reveals systematic gaps. The visualisation is a small-multiples bar chart with the population strata on the x-axis and the metric on the y-axis, plus an annotation showing the absolute gap. Equity dashboards are increasingly mandated by regulators and payers; the BI team owns their integrity.

55.5 Common Pitfalls

What goes wrong

Raw outcome rates without risk-adjustment. The sickest hospital always looks worst; the cleanest comparison is observed vs expected.
PHI in dashboards seen by too many people. Without RLS and small-cell suppression, a dashboard becomes a privacy incident.
Snapshot dashboards in a function that runs in 4-hour ED blocks. Run charts and live wall displays beat daily snapshots.
Composite scores without their inputs. SHMI, HSMR, even patient-satisfaction percentiles — show the components or the audience cannot interrogate the headline.
Inferring causation from administrative data. A correlation in claims data is not a clinical relationship. Caveat clearly.
OR block utilisation reported without case-mix. A surgeon doing complex cases will run lower utilisation than one doing day-cases. Adjust or compare like-for-like.
Equity dashboards that show the gap and not the trend. A 12-percentage-point gap that is closing is a different story from one that is widening; show both.
Care-gap lists distributed by email. Static PDFs go stale within a week. Live dashboards inside the EHR or BI app are the right delivery.
Predicting risk without explaining it. Clinicians do not act on opaque scores. The dashboard must show why a patient is high-risk — the top contributing factors per case.
Forgetting the patient. Provider-facing dashboards rule the BI conversation, but PROMs and patient-portal dashboards close the loop on whether the patient is actually getting better.

55.6 Illustrative Cases

Three case sketches

Yuvijen Health Network mortality funnel plot. Quality team replaces a league-table bar chart of 30-day mortality with a funnel plot of observed vs expected deaths across 14 hospitals. Three hospitals that had ranked at the bottom of the league table fall well within the 95 percent funnel — they took the sickest cases. One hospital that ranked mid-table sits outside the upper 99.8 percent control limit. The board redirects the case-review programme from the apparent worst performer to the actual outlier; mortality at that hospital drops 1.4 points within two quarters after a structured review.

Yuvijen Health bed-occupancy live wall display. Operations team builds a Power BI Phone Layout dashboard that runs on a 65-inch wall display in the bed-management office. Live ward heatmap, ED waiting-time histogram, and a top-10 list of patients ready-for-discharge but awaiting paperwork. Length-of-stay drops 0.7 days within three months — most of the gain comes from the ready-for-discharge list, which had previously been buried in a spreadsheet emailed to the ward at 4 pm.

Yuvijen Health diabetes registry. Population health team builds a Tableau registry dashboard for 38,000 enrolled patients. HbA1c control rate rises from 47 percent to 61 percent within 18 months — not because of new clinical interventions, but because the care-gap list (patients overdue for retinal screening, foot exam, ACE inhibitor) is finally visible to practice managers in real time, with sortable tables that let them work the list down each Friday.

55.7 Hands-On Exercise: Build a Hospital Quality and Operations Dashboard

Three-page hospital dashboard

Aim. Build a three-page healthcare-analytics dashboard in Power BI that ties patient outcomes, operations, and a population-health registry together, with row-level security and small-cell suppression. Tableau equivalents are noted.

Scenario. You are the BI lead at Yuvijen Health Network. The Chief Medical Officer has asked for a dashboard that lets her see, in the Monday quality review, which outcome metrics are out of control, which wards are flowing or blocked, and which patients in the diabetes registry need follow-up this week.

Deliverable. A three-page Power BI report — Outcomes, Operations, Registry — with hospital-level RLS, a small-cell suppression rule of <5 patients per cell, and a CMO summary that consolidates the highest-impact items.

55.7.1 Step 1 — Load and model the data

Use Get Data in Power BI to load five anonymised CSVs:

discharges.csv — DischargeID, PatientHash, AdmitDate, DischargeDate, Hospital, Ward, DRG, Severity, Mortality, Readmit30d.
or_cases.csv — CaseID, Hospital, ORSuite, Surgeon, ScheduledStart, ActualStart, EndTime, CancelFlag, CancelReason.
ed_arrivals.csv — ArrivalID, Hospital, ArrivalTime, TriageTime, DoctorTime, DispositionTime, DepartureTime, Acuity.
bed_occupancy.csv — Hospital, Ward, Hour, OccupiedBeds, TotalBeds.
diabetes_registry.csv — PatientHash, EnrollDate, LastHbA1c, LastHbA1cDate, RetinalLastDate, FootExamLastDate, ACEFlag.

In Power Query, type the columns; never load patient names or identifiers — only PatientHash. Build a DimDate calendar; mark it. Build a DimHospital table with Region, Tier, BedsLicensed. Build a HospitalSecurity table mapping user UPN to allowed hospital codes.

55.7.2 Step 2 — Page 1: Outcomes

Build four visuals.

Mortality funnel plot. Build a calculated table of observed and expected mortality per hospital. Plot a scatter with ExpectedDeaths on x-axis, ObservedRate on y-axis. Use a calculated reference series for the 95 percent and 99.8 percent control limits (computed analytically using the binomial distribution). Conditionally flag hospitals outside the upper 99.8 percent in red.

Readmission cohort curve. Line chart with one line per discharge cohort (monthly), x-axis = days since discharge, y-axis = cumulative readmission percent. Reveals whether readmissions cluster in week 1 or week 3.

Harm-event SPC chart. Line chart with control limits showing weekly counts of hospital-acquired infections, falls, and pressure ulcers. Out-of-control points highlighted.

PROMs paired-line. For a defined cohort (say, post-knee-replacement), pre-op and post-op Oxford Knee Score plotted as paired points joined by a line per patient. Anonymised; aggregated to deciles for views above unit-level access.

DAX measures:

ObservedDeaths =
CALCULATE(SUM(discharges[Mortality]))

ExpectedDeaths =
SUMX(discharges, discharges[ExpectedMortalityProb])

ORatio =
DIVIDE([ObservedDeaths], [ExpectedDeaths])

RiskAdjMortality =
[ORatio] * AVERAGE(network[CrudeMortality])

Tableau alternative: scatter with calculated reference bands; cohort curve native; SPC via reference lines and calculated control limits; paired-line via dual-axis.

55.7.3 Step 3 — Page 2: Operations

Build four visuals.

Bed-occupancy heatmap. Matrix with Ward on rows, Hour on columns, cell colour = percent occupancy. Above 90 percent flagged amber, above 95 percent red. Slicer for Hospital.

Patient-flow Sankey. Power BI Sankey visual showing flow from ED Admit → Ward → Step-Down → Discharge or Transfer. Link width = patient count over the past 7 days.

OR small-multiples. Four panels per OR suite: block utilisation %, FCOTS %, turnover minutes (median), cancellation %. Each as a sparkline trend with reference line at target.

ED time-to histogram. Small-multiples histogram of door-to-triage, door-to-doctor, door-to-decision, disposition-to-departure, with median and 95th percentile markers.

Tableau alternative: heatmap native; Sankey via extension; small-multiples via Trellis; histogram via binned dimension.

55.7.4 Step 4 — Page 3: Diabetes registry

Build three visuals.

Registry control-rate card. Card showing current HbA1c control rate (percent of registered patients with HbA1c < 7 in last 12 months), with trend sparkline below.

Care-gap table. A table sorted by count of overdue items per patient, with columns PatientHash, last HbA1c date, last retinal exam, last foot exam, ACE-on-board flag. Conditional formatting flags overdue items in red. Filtered to patients with 2+ gaps.

HbA1c distribution. Histogram of latest HbA1c across the registry, with reference lines at 7 (good control), 8 (poor), 9 (very poor). Stacked colour by Practice to reveal which practice has the worst tail.

Tableau alternative: card via single-number sheet; table native with conditional formatting; histogram via binned dimension and stacked colour.

55.7.5 Step 5 — Row-Level Security and small-cell suppression

Implement Power BI RLS:

Hospital Manager. Sees their own hospital only.
Network Quality. Sees all hospitals, with PatientHash never displayed in non-summary visuals.
Practice Manager. Sees only their practice’s diabetes-registry slice.
CMO. Sees the network roll-up and the CMO summary page.

Implement small-cell suppression as a DAX measure that returns BLANK when the patient count for a cell is fewer than 5; replace the visual value with the marker Suppressed (n<5). This is a regulatory requirement in most healthcare jurisdictions.

SafePatientCount =
VAR n = CALCULATE(DISTINCTCOUNT(discharges[PatientHash]))
RETURN
IF(n < 5, BLANK(), n)

55.7.6 Step 6 — CMO Monday summary

Build a fourth page that consolidates the highest-priority items each Monday: top three outcome alerts (out-of-control SPC, funnel-plot outliers, readmission cohort breakpoints), top three operational pinch points (over-occupied wards, cancellation spikes, ED 4-hour breaches), top three registry care-gap practices. The page is what the CMO actually opens at the 8 a.m. quality stand-up.

55.7.7 Step 7 — Daily refresh, wall-display layout, and audit logs

Schedule a daily 5 a.m. refresh. Build a wall-display layout (Chapter 47 mobile/responsive principles applied to a fixed large display) with the bed-occupancy heatmap, ED time-to histogram, and OR utilisation tile prominently shown. Configure data-driven alerts for: ward occupancy above 95 percent for 2+ consecutive hours, ED 4-hour breach exceeding 10 percent in any 4-hour block, harm-event SPC out-of-control. Enable workspace audit logs — for healthcare data this is a compliance requirement, not a nice-to-have.

Connect to the Visualisation Layer

Healthcare analytics relies on the visualisation grammar this book has been building, with two specific extensions. Funnel plots and SPC charts (introduced here) are the rendering technique that converts noisy operational data into signal for clinical audiences — the healthcare equivalent of the bullet chart’s role in finance. Heatmaps from Chapter 12 carry bed occupancy. Sankey diagrams (Chapter 49, Chapter 54) carry patient flow. Cohort curves (Chapter 52, Chapter 54) carry readmission. Pareto charts (Chapter 11, Chapter 51) drive the high-risk pyramid. Mobile design (Chapter 47) puts the dashboard on the clinician’s tablet at the bedside. The privacy and RLS discipline of Chapter 36 is more central in healthcare than in any other industry — small-cell suppression is the visualisation-layer expression of patient confidentiality.

Files and Screen Recordings

Power BI three-page hospital dashboard with RLS and small-cell suppression (yuvijen-health-quality.pbix), Tableau equivalent (yuvijen-health-quality.twbx), anonymised workshop dataset (yuvijen-health-data.xlsx), wall-display build (yuvijen-health-wall.pbix), and a screen recording of the dashboard tour (yuvijen-health-walkthrough.mp4) will be embedded here.

Summary

Concept	Description
Healthcare-Dashboard Contract
Patient Privacy Default	PHI behind row-level security, audit logs, suppression of cells with fewer than 5 patients
Risk-Adjustment Mandatory	Raw mortality means nothing without case-mix adjustment
Bedside Reading	Dashboards run on tablets and wall displays at the bedside; speed and font size matter
Three Healthcare Jobs
Patient Outcomes	Are patients getting better, and are we doing harm we could prevent?
Operations	Are wards, theatres, and clinics running well?
Population Health	Which segments of our community need what?
Outcome Families
Mortality	Risk-adjusted in-hospital and 30-day mortality
Readmission	30-day all-cause readmission, condition-specific readmission
Complications and Harm	HACs, hospital-acquired infections, falls, pressure ulcers
Patient-Reported Outcomes	PROMs scores pre and post intervention
Risk-Adjustment Tools
Funnel Plot	Scatter of observed vs expected with 95 and 99.8 percent control limits fanning out
Observed vs Expected	Standard adjustment view that replaces league-table ranking with statistical signalling
Inverted Cohort Curve	Cumulative readmission percent at days 1, 7, 14, 30, 60, 90 reveals concentration shape
SHMI Black Box	Composite hospital mortality measure sensitive to coding depth and palliative exclusions
Operations Visuals
Bed-Occupancy Heatmap	Wards by hour with cell colour as percent occupancy; above 90 is danger zone
Patient-Flow Sankey	Flow from ED admit through ward, ICU, step-down, discharge — bottlenecks visible
Block Utilisation	Actual surgery time over allocated block duration per surgeon
First-Case On-Time Start	Percent of first cases starting within 5 minutes of scheduled
Turnover Time	Minutes between one case ending and the next beginning
Cancellation Rate	Percent of scheduled cases cancelled with reason codes
Door-to-Doctor Clock	Door-to-triage, door-to-doctor, door-to-decision, disposition-to-departure
Run Charts over Snapshots	Run charts and SPC charts beat single-bar snapshots in noisy operational data
Population-Health Tools
Chronic-Disease Registry	Maintained list of patients with a defined condition plus key clinical metrics
Care-Gap List	Sortable list of patients overdue for screening or guideline-recommended care
High-Risk Pyramid	Top 5 percent care management, next 15 percent rising-risk, bottom 80 percent population health
Equity Disparities Lens	Outcome cuts by income, ethnicity, geography, language to surface systematic gaps
Common Pitfalls
Pitfall: Raw Outcome Rates	League tables look bad until risk-adjustment is applied
Pitfall: PHI Without RLS	Patient identifiers in dashboards seen by too many people are a privacy breach
Pitfall: Snapshot Dashboards	Snapshot views miss the trend that healthcare audiences need to see
Pitfall: Composite Without Inputs	SHMI and similar without coding-depth context invite misreading
Pitfall: Causal Claims from Claims	Correlations in claims data are not clinical relationships
Pitfall: Block Utilisation Without Mix	Surgeons doing complex cases run lower utilisation than day-case surgeons
Pitfall: Equity Without Trend	Equity dashboards must show whether the gap is narrowing or widening
Pitfall: Care-Gap PDF Distribution	Static PDFs go stale within a week; live dashboards close the loop
Pitfall: Opaque Risk Scores	Clinicians do not act on risk scores without contributing-factor explanations
Pitfall: Forgetting the Patient	Provider-facing dashboards must close the loop with PROMs and patient-portal views
Hands-On Dashboard
Page 1 — Outcomes	Funnel plot, readmission cohort curve, harm-event SPC, PROMs paired-line
Page 2 — Operations	Bed-occupancy heatmap, patient-flow Sankey, OR small-multiples, ED time-to histograms
Page 3 — Registry	Registry control-rate card, care-gap table, HbA1c distribution
Hospital RLS	Hospital Manager, Network Quality, Practice Manager, CMO roles with separate scopes
Small-Cell Suppression	DAX measure returns BLANK when patient count below 5; replaces value with Suppressed marker
CMO Monday Summary	Top three outcome alerts, three operational pinch points, three registry gaps consolidated
Wall-Display Layout	Bed occupancy, ED time-to histogram, OR tile on 65-inch wall display in bed-management office