11.2 Selecting Appropriate Healthcare Visualizations

Selecting Appropriate Healthcare Visualizations

Selecting the correct visualization is a foundational competency for the Certified Health Data Analyst (CHDA). In clinical and operational analytics, choosing the wrong chart type does not merely create an aesthetic failure; it can obscure quality signals, misrepresent hospital performance, lead to erroneous clinical conclusions, or trigger inappropriate operational interventions. A health data analyst must systematically match the analytical objective, underlying data structure, distribution properties, and audience requirements to the optimal graphical representation.

---

1. Comprehensive Healthcare Visualization Taxonomy

Healthcare analytical inquiries generally fall into six primary analytical objectives. Mapping the business or clinical question to its core objective dictates the allowable family of visualizations.

+---------------------------------------------------------------------------------------------------+
|                               HEALTHCARE VISUALIZATION OBJECTIVE TAXONOMY                         |
+-------------------+--------------------+--------------------+-------------------------------------+
| 1. COMPOSITION    | 2. COMPARISON      | 3. DISTRIBUTION    | 4. RELATIONSHIP / CORRELATION       |
| - 100% Stacked Bar| - Horizontal Bar   | - Histogram        | - Scatter Plot (2 continuous vars)  |
| - Treemap (Hier.) | - Grouped Column   | - Box-and-Whisker  | - Bubble Chart (3 continuous vars)  |
| - Waterfall Chart | - Slopegraph       | - Violin Plot (KDE)| - Correlation Heatmap Matrix        |
+-------------------+--------------------+--------------------+-------------------------------------+
| 5. PROCESS & PATIENT FLOW              | 6. TEMPORAL EVOLUTION & LONGITUDINAL TRENDS              |
| - Sankey Diagram (Acuity Transitions)  | - Run Chart (Median Centerline, Non-parametric Rules)    |
| - Alluvial Flow Diagram                | - Statistical Process Control (SPC) Shewhart Charts      |
| - Swimlane Workflow Process Map        |   (Mean Centerline, ± 3-Sigma Control Limits)            |
+----------------------------------------+----------------------------------------------------------+

Breakdown of Analytical Objectives:

1. Composition (Part-to-Whole): Illustrating how individual sub-components constitute a complete total. Examples include hospital payer mix breakdown (Medicare, Medicaid, Commercial, Self-Pay) or total operating expense partitioning across clinical departments. Recommended charts: 100% Stacked Bar Charts, Treemaps (for multi-level hierarchical categorical data), and Waterfall Charts (for step-by-step financial variance bridges).
2. Comparison (Discrete Categories): Evaluating magnitudes across distinct, categorical entities. Examples include comparing 30-day readmission rates across 15 hospital service lines or average ED wait times across 6 satellite clinics. Recommended charts: Horizontal Bar Charts (sorted in descending order), Grouped Column Charts, and Slopegraphs (comparing pre- vs. post-intervention periods).
3. Distribution (Shape, Spread & Outliers): Examining the statistical spread, central tendency, skewness, and extreme values of continuous clinical variables. Examples include Inpatient Length of Stay (LOS), ICU drug acquisition costs, or Door-to-Needle times. Recommended charts: Histograms, Box-and-Whisker Plots, and Violin Plots.
4. Relationship & Correlation: Identifying associations, dependencies, clusters, or causal patterns between two or more continuous variables. Examples include evaluating the correlation between nurse staffing ratios and hospital-acquired pressure injury rates, or patient age vs. total surgical episode cost. Recommended charts: Scatter Plots and Bubble Charts.
5. Process & Patient Flow: Mapping the sequential, non-linear progression of patients or clinical resources through healthcare delivery stages. Examples include tracking patient transitions from ED triage through observation, inpatient admission, surgical intervention, and post-acute discharge destinations. Recommended charts: Sankey Diagrams and Swimlane Process Maps.
6. Temporal Evolution & Longitudinal Quality: Monitoring clinical indicators over time to distinguish true process improvements from random noise. Examples include tracking monthly central-line infection rates or daily surgical suite turnover times. Recommended charts: Run Charts and Statistical Process Control (SPC) Charts.

---

2. Longitudinal Quality Improvement: Run Charts vs. Statistical Process Control (SPC)

In healthcare quality improvement (QI), understanding variation is paramount. Walter A. Shewhart and W. Edwards Deming established that all processes exhibit variation, categorized into:

• Common Cause Variation: The natural, inherent, predictable background noise of a stable system. Attempting to react to every individual data point in common cause variation is termed "tampering", which increases overall process instability and cost.
• Special Cause Variation: Unnatural, assignable-cause variation resulting from an external disturbance, process shift, or deliberate quality improvement intervention. Special cause variation signals a fundamental change in the system that warrants immediate root-cause investigation or validation of improvement.

+---------------------------------------------------------------------------------------------------+
|                                RUN CHARTS VS. SPC CONTROL CHARTS                                  |
+-------------------------------------------------+-------------------------------------------------+
| RUN CHART                                       | STATISTICAL PROCESS CONTROL (SPC) CHART         |
| - Centerline: Historical MEDIAN                 | - Centerline: Historical MEAN                   |
| - Dispersion Limits: NONE                       | - Dispersion Limits: Upper/Lower Control Limits |
| - Minimum Data Points: n >= 10-12               |   at exactly ± 3-Sigma (UCL and LCL)            |
| - Statistical Rules: Non-parametric             | - Minimum Data Points: n >= 20-30               |
|   (Shift, Trend, Runs, Astronomical Outlier)    | - Statistical Rules: Nelson / Western Electric  |
| - Best For: Initial QI pilots, rapid feedback   | - Best For: Formal QI, regulatory surveillance  |
+-------------------------------------------------+-------------------------------------------------+

Run Charts: Principles & Interpretation Rules

A Run Chart is a line chart displaying data over time with a central reference line representing the historical Median. Run charts do not require complex distributional assumptions, making them ideal for initial quality improvement initiatives when data points are limited ($n \ge 10\text{ to }12$).

The Four Standard Run Chart Rules (Perla / IHI Standard):

1. Shift: A run of 6 or more consecutive data points falling entirely above or entirely below the median. Data points falling exactly on the median line do not count and are omitted from the calculation.
2. Trend: A continuous sequence of 5 or more consecutive data points that are continually increasing or continually decreasing. Ties between consecutive points do not count as increases or decreases.
3. Runs Test (Too Few or Too Many Runs): A "run" is a continuous sequence of consecutive data points on one side of the median. The total number of runs across the dataset is counted and compared against statistical reference tables based on sample size $N$ (excluding points on the median). Having fewer runs than expected indicates non-random grouping or process shift; having more runs than expected indicates negative autocorrelation or data oscillation.
4. Astronomical Data Point: An extraordinarily extreme outlier that is immediately obvious to any observer without statistical calculation. It represents a dramatic, blatant special cause event.

Statistical Process Control (SPC) Shewhart Charts

An SPC Control Chart is an advanced time-series tool featuring a central line representing the historical Mean ($\bar{X}, \bar{p}, \bar{u}$), bounded by an Upper Control Limit (UCL) at $+3\sigma$ and a Lower Control Limit (LCL) at $-3\sigma$, derived from the process standard deviation.

$\text{Upper Control Limit (UCL)} = \text{Process Mean} + 3\sigma$ $\text{Lower Control Limit (LCL)} = \max(0, \text{Process Mean} - 3\sigma)$

Control limits represent the mathematical voice of the process (what the process is currently delivering), whereas clinical specification limits represent the voice of the customer/regulator (what is desired). Control limits must never be confused with clinical targets or specification thresholds.

SPC Zone Architecture (Three-Sigma Partitions):

• Zone C: Centerline to $\pm 1\sigma$ (covers $68.27%$ of normal variation)
• Zone B: $\pm 1\sigma$ to $\pm 2\sigma$ (covers $27.18%$ of normal variation)
• Zone A: $\pm 2\sigma$ to $\pm 3\sigma$ (covers $4.28%$ of normal variation)

  +3 Sigma ──────────────────────────────────────────────────────── UCL
               ZONE A  (2 to 3 Sigma)
  +2 Sigma ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─
               ZONE B  (1 to 2 Sigma)
  +1 Sigma ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─
               ZONE C  (0 to 1 Sigma)
    Mean   ──────────────────────────────────────────────────────── CENTERLINE
               ZONE C  (0 to 1 Sigma)
  -1 Sigma ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─
               ZONE B  (1 to 2 Sigma)
  -2 Sigma ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─
               ZONE A  (2 to 3 Sigma)
  -3 Sigma ──────────────────────────────────────────────────────── LCL

The Nelson / Western Electric Rules for Special Cause Variation in Healthcare SPC:

1. Rule 1 (Outlier Beyond Control Limits): One point falls beyond Zone A ($> 3\sigma$ from the centerline, beyond UCL or LCL). Probability under normal variation is $P = 0.0027$ ($0.27%$). Indicates an acute special cause event.
2. Rule 2 (Zone C Shift): Nine (or 8) consecutive points fall on the same side of the centerline (all in Zone C or beyond). Indicates a persistent process level shift.
3. Rule 3 (Trend): Six consecutive points continually increase or continually decrease. Indicates a continuous drift or sustained wear/improvement.
4. Rule 4 (Sawtooth Oscillation): Fourteen consecutive points alternate strictly up and down. Indicates over-control, negative autocorrelation, or alternating data sources (e.g., day shift vs. night shift).
5. Rule 5 (Zone A Clustering): Two out of three consecutive points fall in Zone A ($> 2\sigma$) on the same side of the centerline.
6. Rule 6 (Zone B Clustering): Four out of five consecutive points fall in Zone B or Zone A ($> 1\sigma$) on the same side of the centerline.
7. Rule 7 (Stratification / Hugging Centerline): Fifteen consecutive points fall within Zone C (within $\pm 1\sigma$ of the centerline). Indicates that the process standard deviation was miscalculated or data from multiple distinct sub-processes were inadvertently combined.
8. Rule 8 (Mixture Pattern): Eight consecutive points fall on both sides of the centerline with zero points falling within Zone C. Indicates a mixture of two distinct underlying populations (e.g., combining data from two separate surgical operating rooms with different baseline speeds).

---

3. SPC Control Chart Selection Algorithm & Mathematical Formulations

Selecting the correct SPC control chart is strictly governed by whether the data is Continuous (Variables data) or Discrete (Attribute data), the subgroup sample size, and whether the inspection unit is constant or variable.

+---------------------------------------------------------------------------------------------------+
|                                 SPC CONTROL CHART SELECTION DECISION ENGINE                       |
+---------------------------------------------------------------------------------------------------+
                                                  │
                                  ┌───────────────┴───────────────┐
                                  ▼                               ▼
                          [CONTINUOUS DATA]               [DISCRETE ATTRIBUTE]
                                  │                               │
                  ┌───────────────┴───────────────┐               │
                  ▼                               ▼               │
             [Subgroup n = 1]            [Subgroup n >= 2]        │
                  │                               │               │
               [I-MR]                     ┌───────┴───────┐       │
             (Indiv / Moving Range)       ▼               ▼       │
                                      [2 <= n <= 9]    [n >= 10]  │
                                         [X-bar & R]  [X-bar & S] │
                                                                  │
          ┌───────────────────────────────────────────────────────┴───────────────────────────────┐
          ▼                                                                                       ▼
  [DEFECTIVE UNITS / PROPORTIONS]                                                 [COUNT OF DEFECTS / RATES]
  (Classification: Yes/No, Defective/Non-defective)                              (Occurrences per Area of Opportunity)
          │                                                                                       │
  ┌───────┴───────┐                                                                       ┌───────┴───────┐
  ▼               ▼                                                                       ▼               ▼
[Variable n_i]  [Constant n]                                                            [Variable Area n_i]  [Constant Unit]
  [p-Chart]       [np-Chart]                                                              [u-Chart]             [c-Chart]

Continuous (Variables) Data Control Charts

1. Individual and Moving Range ($I\text{-}MR$ / $X\text{-}mR$) Chart

• Criteria: Used when continuous data is collected as single individual observations per time period ($n=1$).
• Healthcare Applications: Monthly hospital-wide Average Length of Stay (ALOS), daily surgical supply expenditure, weekly pharmacy budget variance.
• Mathematical Formulas:
  • Moving Range between consecutive points: $MR_i = |X_i - X_{i-1}|$
  • Mean of observations: $\bar{X} = \frac{\sum X_i}{k}$, Mean Moving Range: $\bar{MR} = \frac{\sum MR_i}{k-1}$
  • Individuals ($I$) Chart Limits: $UCL = \bar{X} + 2.66 \bar{MR}$, $LCL = \bar{X} - 2.66 \bar{MR}$
  • Moving Range ($MR$) Chart Limits: $UCL_{MR} = 3.267 \bar{MR}$, $LCL_{MR} = 0$

2. $\bar{X}$ and $R$ (X-bar and Range) Chart

• Criteria: Used for continuous data collected in small, rational subgroups of size $2 \le n \le 9$.
• Healthcare Applications: Sampling 5 laboratory turnaround times per 8-hour shift; measuring blood pressure reduction across cohorts of 6 clinical trial participants.
• Mathematical Formulas:
  • Subgroup Mean: $\bar{X}i = \frac{\sum X}{n}$, Subgroup Range: $R_i = X{\max} - X_{\min}$
  • Grand Mean: $\bar{\bar{X}}$, Average Range: $\bar{R}$
  • $\bar{X}$ Chart Limits: $UCL = \bar{\bar{X}} + A_2 \bar{R}$, $LCL = \bar{\bar{X}} - A_2 \bar{R}$ (where $A_2$ is a standard Shewhart factor, e.g., $A_2 = 0.577$ for $n=5$)
  • $R$ Chart Limits: $UCL_R = D_4 \bar{R}$, $LCL_R = D_3 \bar{R}$

3. $\bar{X}$ and $S$ (X-bar and Standard Deviation) Chart

• Criteria: Used for continuous data collected in larger subgroups of size $n \ge 10$, where the sample standard deviation $s$ provides a more statistically efficient estimate of dispersion than the range $R$.
• Healthcare Applications: High-volume automated clinical chemistry analyzer calibration batches ($n=20$).

---

Discrete (Attribute) Data Control Charts

1. $p$-Chart (Proportion Defective, Variable Subgroup Size)

• Criteria: Measures the proportion of non-conforming (defective) items where each unit is evaluated binarily (e.g., compliant vs. non-compliant, readmitted vs. not readmitted), and the denominator (subgroup size $n_i$) varies across time periods.
• Healthcare Applications: Monthly 30-day all-cause hospital readmission rate ($n_i = \text{monthly discharges}$), surgical site infection proportion, percentage of incomplete clinical documentation records.
• Mathematical Formulas: $\text{Subgroup Proportion: } p_i = \frac{x_i}{n_i}, \quad \text{Centerline: } \bar{p} = \frac{\sum x_i}{\sum n_i}$ $\text{Control Limits for Period } i: \quad UCL_i = \bar{p} + 3\sqrt{\frac{\bar{p}(1-\bar{p})}{n_i}}, \quad LCL_i = \max\left(0, \bar{p} - 3\sqrt{\frac{\bar{p}(1-\bar{p})}{n_i}}\right)$ Note: Because $n_i$ fluctuates, the control limits widen in low-volume months and narrow in high-volume months, producing characteristic "jagged" or step-like control limits.

2. $np$-Chart (Number of Defectives, Constant Subgroup Size)

• Criteria: Measures the raw count of defective units when the subgroup sample size $n$ remains strictly constant across all observation periods.
• Healthcare Applications: Auditing coding error counts in fixed weekly random samples of exactly $n = 100$ inpatient discharge charts.
• Mathematical Formulas: $\text{Centerline: } n\bar{p}, \quad UCL = n\bar{p} + 3\sqrt{n\bar{p}(1-\bar{p})}, \quad LCL = \max\left(0, n\bar{p} - 3\sqrt{n\bar{p}(1-\bar{p})}\right)$

3. $u$-Chart (Rate of Defects per Unit, Variable Inspection Area)

• Criteria: Measures the rate of defects or events that can occur multiple times per unit of exposure (Poisson distribution), where the area of opportunity (exposure denominator $n_i$) varies across periods.
• Healthcare Applications: Central Line-Associated Bloodstream Infections (CLABSI) per 1,000 central line days; Catheter-Associated Urinary Tract Infections (CAUTI) per 1,000 catheter days; inpatient falls per 1,000 patient bed days across hospital units with varying monthly bed occupancy.
• Mathematical Formulas: $\text{Subgroup Rate: } u_i = \frac{c_i}{n_i}, \quad \text{Centerline: } \bar{u} = \frac{\sum c_i}{\sum n_i}$ $\text{Control Limits for Period } i: \quad UCL_i = \bar{u} + 3\sqrt{\frac{\bar{u}}{n_i}}, \quad LCL_i = \max\left(0, \bar{u} - 3\sqrt{\frac{\bar{u}}{n_i}}\right)$

4. $c$-Chart (Count of Defects, Constant Inspection Unit)

• Criteria: Measures the total count of defects occurring on a single, fixed, constant inspection unit ($n=1$ constant unit of opportunity).
• Healthcare Applications: Total medication administration errors recorded in a single specific 24-bed ICU per calendar month; total non-compliance infractions found during monthly facility physical safety inspections.
• Mathematical Formulas: $\text{Centerline: } \bar{c} = \frac{\sum c_i}{k}, \quad UCL = \bar{c} + 3\sqrt{\bar{c}}, \quad LCL = \max\left(0, \bar{c} - 3\sqrt{\bar{c}}\right)$

---

4. Specialized Healthcare Visualizations

Funnel Plots for Institutional Performance Benchmarking

In healthcare performance benchmarking (e.g., CMS Hospital Compare, surgeon scorecards), comparing raw mortality or complication rates across institutions using simple ranking tables or "league tables" introduces severe sample size bias. Small community hospitals with low procedural volume naturally exhibit extreme high and low rates purely due to random binomial variation, unfairly penalizing or rewarding them.

• Architecture:
  • X-axis: Institutional volume or exposure denominator (e.g., total annual CABG procedures performed).
  • Y-axis: Observed performance metric (e.g., 30-day risk-adjusted mortality rate, or Observed/Expected [O/E] mortality ratio).
  • Centerline: National or statewide benchmark rate.
  • Funnel Limits: Statistical control limits based on binomial or Poisson exact distribution limits that form a funnel shape: wide at low volumes (reflecting high expected sampling variance) and narrow at high volumes.
  • Inner Limits (95% CI / $\pm 2\text{SD}$): "Alert limits" (warning threshold, $P < 0.05$).
  • Outer Limits (99.8% CI / $\pm 3\text{SD}$): "Action limits" (alarm threshold, $P < 0.002$).
• Interpretation: Institutions plotting inside the funnel limits are performing within expected statistical variation (common cause). Institutions plotting above the upper 99.8% limit are true negative outliers requiring immediate regulatory intervention; institutions plotting below the lower 99.8% limit represent superior benchmark performers.

  Observed
  Mortality
    ▲
    │  * (Small Vol Outlier)                    Upper 99.8% Limit (Action)
    │   \                                      / ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─
    │    \    *                               /  Upper 95.0% Limit (Alert)
    │     \ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─/ ─────────────────────────
    │══════\════════════════════════════════/═════ National Benchmark (Mean)
    │       \ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─/ ───────────────────────────
    │        \                            /   Lower 95.0% Limit
    │         \                          / ─ ─ Lower 99.8% Limit
    │          *                        *
    └──────────────────────────────────────────────────────────────────►
    Low Institutional Volume                     High Institutional Volume

Box-and-Whisker & Violin Plots for Skewed Clinical Distributions

Continuous clinical and financial healthcare variables (such as Inpatient Length of Stay [LOS], Emergency Department wait times, and total cost per discharge) are heavily right-skewed and non-normal. Parametric metrics like mean and standard deviation distort these distributions. Box-and-whisker and violin plots provide robust non-parametric representations.

• Tukey Box-and-Whisker Plot Anatomy:
  • Median ($Q_2$ / 50th percentile): Central line inside the box.
  • Interquartile Range ($IQR$): Bounded by the 25th percentile ($Q_1$) and 75th percentile ($Q_3$). The box height equals $IQR = Q_3 - Q_1$, containing the central 50% of patient encounters.
  • Whiskers: Extend from $Q_1$ down to the smallest data point within $Q_1 - 1.5 \times IQR$, and from $Q_3$ up to the largest data point within $Q_3 + 1.5 \times IQR$.
  • Outliers: Individual points plotted beyond the whiskers ($> 1.5 \times IQR$ from the box hinges), representing extreme clinical outliers.
• Violin Plots: Combine a box plot with a mirrored Kernel Density Estimation (KDE) plot. Violin plots reveal multi-modal distributions (e.g., bimodal surgical lengths of stay representing uncomplicated day-stay patients vs. patients suffering postoperative complications) that are completely hidden by standard box plots.

Heatmaps for Clinical and Operational Density

A Heatmap represents a 2D matrix where individual cell values are encoded by color saturation. In healthcare operations, heatmaps identify temporal and spatial bottlenecks:

• ED Patient Arrival Density: Mapping hour of the day (00:00 to 23:00 on y-axis) against day of the week (Monday through Sunday on x-axis). Reveals predictable arrival surges (e.g., Monday 10:00 AM – 2:00 PM), allowing clinical leadership to optimize physician and nursing shift scheduling.
• Hospital Unit Bed Occupancy: Displaying real-time bed occupancy percentages across 20 acute care nursing units over the preceding 30 days.

Sankey Diagrams for Care Continuum Pathways

A Sankey Diagram is a directed flow visualization where the width of connecting bands is strictly proportional to the volume of flow. In healthcare analytics, Sankey diagrams map complex, non-linear patient journeys across the care continuum:

• Clinical Pathway Flow: Visualizing patient progression from ED Triage (ESI Levels 1–5) $\rightarrow$ Initial Clinical Placement (Direct Discharge, Clinical Decision Observation Unit, Inpatient Floor) $\rightarrow$ Level of Care Escalation (Step-Down Unit, Intensive Care Unit) $\rightarrow$ Final Discharge Destination (Home Self-Care, Home Health with Skilled Nursing, Skilled Nursing Facility [SNF], Long-Term Acute Care [LTACH], Hospice, or Expired).

Forest Plots for Clinical Meta-Analyses and Subgroup Odds Ratios

A Forest Plot (blobbogram) is the gold-standard graphical method for synthesizing evidence from multiple randomized controlled clinical trials or epidemiological subgroup analyses:

• Vertical Line of No Effect: Positioned at $\text{Odds Ratio (OR)} = 1.0$, $\text{Relative Risk (RR)} = 1.0$, or $\text{Mean Difference} = 0.0$.
• Study Point Estimates: Plotted as square markers whose area is proportional to the statistical weight (sample size / precision) of that individual study.
• Confidence Interval Whiskers: Horizontal lines extending from each square representing the 95% confidence interval of the effect estimate for that study.
• Diamond Summary Estimate: Positioned at the bottom of the forest plot; the center of the diamond represents the pooled meta-analytic effect estimate, and the horizontal width of the diamond represents the 95% confidence interval of the pooled result.

---

5. Master Chart Selection Decision Matrix Table