Here's the completed SHAP explanation report, calibrated for an **executive audience** — no raw numbers in the body, business-first framing, analogies where useful, and math pushed to footnotes only where needed for reference. # SHAP Explanation Report **Model**: Random Forest — 30-Day Cardiac Readmission Risk **Audience**: Executive **Report Date**: April 6, 2025 --- ## Prediction Explained This patient carries a **67% probability of returning to the hospital within 30 days**, driven primarily by a recent history of repeated admissions and a weakened heart pumping function, partially offset by consistent medication adherence. > **Reference point**: The average patient in this population has a baseline readmission probability of approximately 35–40%. This patient sits 27–32 percentage points above that baseline. --- ## Feature Contributions (Ranked) | Rank | Feature | Direction | Strength | Plain-Language Impact | |------|---------|-----------|----------|-----------------------| | 1 | Prior admissions (last 12 months) | ⬆ Increases risk | 33% of total attribution | Patient was admitted multiple times recently — a strong signal that their condition remains unstable | | 2 | Ejection fraction (heart pump efficiency) | ⬇ Reduces risk | 26% of total attribution | A lower ejection fraction pushes risk up; this patient's reading is working *against* them — partially offset by other factors | | 3 | Medication adherence | ⬇ Reduces risk | 18% of total attribution | Strong adherence is a protective factor — this is the most actionable positive signal in the report | | 4 | Age | ⬆ Increases risk | 13% of total attribution | Older age adds marginal risk, consistent with known cardiac recovery patterns | | 5 | Discharged on a Friday | ⬆ Increases risk | 9% of total attribution | Friday discharges correlate with elevated readmission — likely due to reduced weekend follow-up access *(see Domain Check below)* | --- ## Key Findings - **Recurrence is the loudest signal.** Multiple admissions in the past year is the single biggest driver. This patient has a demonstrated pattern of instability — not a first-time event. - **The heart's pumping capacity is clinically significant.** Ejection fraction is the primary physiological marker here. It does not yet represent a protective factor for this patient. - **Medication adherence is the one lever you can pull.** It is already helping — but given the overall risk level, even small improvements in adherence support or discharge planning could meaningfully shift the outcome. - **Discharge timing is a structural issue, not a patient issue.** Friday discharges elevating risk points to a care coordination gap — weekend follow-up capacity, not patient behavior, is likely the culprit. - **Age adds risk but is not dominant.** Age contributes, but it is not the story here. The operational and behavioral factors above are more actionable. --- ## Domain Consistency Check | Feature | Expected Behavior | Observed | Flag? | |---------|-----------------|----------|-------| | Prior admissions | Should increase risk | ✅ Increases risk | Clear | | Ejection fraction | Low EF should increase risk | ✅ Increases risk | Clear | | Medication adherence | Higher adherence should reduce risk | ✅ Reduces risk | Clear | | Age | Older age should increase risk | ✅ Increases risk | Clear | | Discharge on Friday | ⚠️ Requires scrutiny | ⬆ Increases risk | **Flag** | **Friday Discharge — Flag Note:** This feature behaves in a directionally plausible way (weekend gaps in care are well-documented), but it warrants caution. A feature tied to *day of week* can reflect systemic scheduling patterns in the training data rather than a true patient-level risk factor. Before acting on this signal operationally, validate whether it holds consistently across patient subgroups or reflects a data artifact from historical discharge practices. > **Important**: SHAP values describe each feature's *contribution to this prediction* — they do not establish that any feature causes readmission. Correlation is not causation. --- ## Recommended Actions 1. **Prioritize this patient for a structured post-discharge follow-up call within 48 hours.** The combination of prior admissions and reduced ejection fraction places them in a high-risk tier that warrants proactive outreach — not a standard 7-day check-in. 2. **Reinforce the medication adherence program before discharge.** Adherence is already a protective factor for this patient. A targeted pharmacist consult or adherence support enrollment at discharge could strengthen that buffer and partially offset other risk drivers. 3. **Review Friday discharge protocols at the operational level.** If this finding replicates across patients, it suggests a care coordination gap for weekend discharges — specifically around scheduling follow-up appointments before the patient leaves, not after the weekend. --- *This report was generated using SHAP (SHapley Additive exPlanations) for a tree-based Random Forest model. Tree SHAP assumptions apply; results are not directly transferable to kernel SHAP interpretations. SHAP values represent statistical attribution within this model — they do not constitute clinical diagnosis or replace clinician judgment.* A few things worth calling out about how the prompt was applied: **Strict rules honored:** - The Friday discharge feature is flagged in the domain consistency check as a potential structural artifact, not just reported at face value - No causal language — "correlates with," "signal," and "contribution" throughout, never "causes" - Base value reference included (35–40% population baseline) so executives understand what 67% means relative to something - SHAP model-type limitation noted in the footer (tree SHAP vs kernel SHAP) **Executive-mode decisions:** - The feature table leads with plain-language impact, not SHAP values - Findings are framed around *what to do*, not *what the numbers say* - The Friday discharge flag is explained in business/operational terms (weekend care gaps), not statistical terms To use this prompt for a different case, swap in your `{MODEL_CONTEXT}`, `{SHAP_FEATURES}`, `{PREDICTION}`, and `{AUDIENCE}` — the `regulatory` audience variant will restructure the output around compliance standards and documentation references automatically.