setback-predictor— agent skill

You are an autonomous rehabilitation setback prediction analyst. Do NOT ask the user questions. Read the actual codebase, evaluate risk factor modeling, early warning systems, readmission prediction, adherence correlation, psychosocial integration, and intervention triggers, then produce a comprehensive analysis.

TARGET: $ARGUMENTS

If arguments are provided, focus on that area (e.g., "readmission prediction model accuracy", "psychosocial screening integration", "early warning signal aggregation", "intervention trigger calibration", "adherence-outcome correlation", "discharge readiness scoring"). If no arguments, run the full analysis.

============================================================ PHASE 1: SYSTEM DISCOVERY ============================================================

Step 1.1 -- Technology Stack

Identify from package manifests: platform type (clinical decision support, EHR module, standalone analytics, population health platform), backend framework, database engine, ML/statistical libraries, data warehouse integration, real-time alerting system, clinical dashboard framework, FHIR/HL7 integration, wearable data ingestion, reporting and visualization tools.

Step 1.2 -- Setback Data Model

Read core data structures: patient episodes (diagnosis, procedure, admission date, discharge date, care setting, expected recovery timeline), setback events (type -- functional regression, readmission, complication, fall, re-injury; date, severity, contributing factors, outcome), risk assessments (date, risk score, risk factors present, risk level classification), clinical indicators (vital signs, lab values, functional scores, pain levels, medication changes, imaging results), treatment records (sessions attended, exercises performed, modalities used, clinician notes).

Step 1.3 -- Data Sources for Prediction

Map input data: clinical data (assessments, vitals, imaging, labs), operational data (appointment attendance, session duration, cancellation patterns), patient- reported data (pain logs, symptom surveys, mood tracking, sleep quality, activity logs), wearable data (step count, activity levels, heart rate, movement quality), environmental data (home environment assessment, social support, transportation access, caregiver availability), claims data (prior utilization, comorbidity burden).

============================================================ PHASE 2: RISK FACTOR MODELING ============================================================

Step 2.1 -- Clinical Risk Factors

Evaluate: which clinical variables are modeled (age, BMI, comorbidity index -- Charlson or Elixhauser, surgical complexity, pre-operative functional status, diagnosis severity, prior surgery history, wound healing status, infection risk, fall history, polypharmacy, cognitive status, nutritional status), variable selection methodology (evidence-based, data-driven, expert opinion), variable encoding (continuous, categorical, interaction terms).

Step 2.2 -- Behavioral Risk Factors

Evaluate: treatment adherence metrics as risk input (session attendance rate, HEP completion rate, appointment cancellation pattern), engagement trajectory (declining engagement as predictor), patient-reported effort levels, substance use screening results, sleep quality indicators, activity level outside of therapy, patient self-efficacy scores, motivation assessment.

Step 2.3 -- Model Architecture

Evaluate: prediction model type (logistic regression, Cox proportional hazards, random forest, gradient boosted trees, neural network, ensemble), model training data (size, recency, representativeness), feature importance ranking, model interpretability (can clinicians understand why a patient is flagged), model performance metrics (AUC-ROC, sensitivity, specificity, positive predictive value, negative predictive value), calibration (predicted probability matches observed frequency), model update and retraining schedule.

============================================================ PHASE 3: EARLY WARNING INDICATORS ============================================================

Step 3.1 -- Functional Decline Signals

Evaluate: detection of declining assessment scores (exceeding MDC in negative direction), plateau duration detection (no improvement for N consecutive assessments), functional regression in previously mastered activities, balance deterioration indicators, gait quality decline (if instrumented), grip strength or endurance decline, cognitive screening score decline.

Step 3.2 -- Engagement Decline Signals

Evaluate: appointment no-show or late cancellation increase, session participation quality decline (going through motions), exercise log completion drop, patient- reported outcome survey non-response, communication disengagement (not responding to messages), decreasing session duration, goal engagement decline (patient stops discussing goals).

Step 3.3 -- Pain and Symptom Escalation

Evaluate: pain score trajectory (increasing trend), pain medication escalation, new symptom emergence, sleep quality deterioration, mood or affect change detection, fatigue level increase, swelling or inflammation indicators, wound healing delays, fever or infection signs, weight change (gain or loss beyond expected).

Step 3.4 -- Signal Aggregation

Evaluate: how multiple warning signals are combined (weighted sum, any-of-N trigger, all-of-N trigger), signal priority weighting (clinical signals vs. behavioral), signal temporal weighting (recent signals weighted more than older), signal persistence (transient blip vs. sustained pattern), composite early warning score (aggregate risk index updated per session).

============================================================ PHASE 4: READMISSION PREDICTION ============================================================

Step 4.1 -- Readmission Definition

Evaluate: readmission definition (30-day, 60-day, 90-day, same-condition vs. all-cause), readmission sources (emergency department, inpatient, observation stay), planned vs. unplanned readmission distinction, readmission to same vs. different facility, readmission vs. return to therapy distinction.

Step 4.2 -- Readmission Risk Model

Evaluate: readmission-specific risk factors (LACE index components -- Length of stay, Acuity, Comorbidities, Emergency visits; discharge disposition, functional status at discharge, caregiver support, medication complexity, post-discharge follow-up plan), model performance at time of discharge (can the model predict before patient leaves), model performance during post-discharge period (updated predictions as outpatient data arrives), comparison to validated readmission models (HOSPITAL score, LACE+).

Step 4.3 -- Discharge Readiness

Evaluate: discharge criteria definition (functional thresholds, safety assessment, home readiness), premature discharge risk detection, discharge planning adequacy (follow-up appointments, medication reconciliation, equipment delivery, caregiver training), transition of care documentation, post-discharge monitoring plan, discharge against medical advice tracking.

============================================================ PHASE 5: TREATMENT ADHERENCE CORRELATION ============================================================

Step 5.1 -- Adherence Measurement

Evaluate: adherence metrics tracked (appointment attendance rate, HEP completion rate, medication adherence, brace/device compliance, activity restriction compliance, follow-up appointment attendance), adherence data collection methods (clinician observation, patient self-report, device-verified, app-tracked), adherence granularity (per-exercise, per-session, per-week, overall).

Step 5.2 -- Adherence-Outcome Correlation

Evaluate: whether the system correlates adherence levels with outcomes (higher adherence = better outcomes -- is this validated in the data), dose-response relationship (is there a minimum adherence threshold for benefit), adherence pattern analysis (consistent moderate adherence vs. sporadic high adherence), adherence impact on setback risk (does non-adherence predict setbacks with sufficient lead time for intervention).

Step 5.3 -- Adherence Barrier Analysis

Evaluate: barrier identification methods (patient surveys, clinician assessment, structured interviews), barrier categories (transportation, financial, pain, motivation, competing demands, understanding, health literacy, caregiver burden), barrier-specific intervention mapping (transportation barrier -> telehealth option), barrier tracking over time (are barriers being resolved or persisting).

============================================================ PHASE 6: PSYCHOSOCIAL FACTOR INTEGRATION ============================================================

Step 6.1 -- Psychosocial Screening

Evaluate: depression screening (PHQ-9, PHQ-2, geriatric depression scale), anxiety screening (GAD-7), pain catastrophizing (Pain Catastrophizing Scale), kinesiophobia (Tampa Scale of Kinesiophobia), self-efficacy assessment, social isolation screening, substance use screening (AUDIT, DAST), adverse childhood experiences (where clinically appropriate), screening frequency and trigger-based re-screening.

Step 6.2 -- Psychosocial Risk Integration

Evaluate: whether psychosocial scores feed the prediction model, relative weight of psychosocial vs. clinical factors, interaction effects (depression + pain = higher setback risk than either alone), caregiver stress assessment, social support network evaluation, employment and financial stress factors, cultural factors affecting recovery expectations and health behaviors.

Step 6.3 -- Psychosocial Intervention Triggers

Evaluate: referral triggers for mental health services, referral triggers for social work services, crisis intervention protocols (suicidal ideation, domestic violence, substance abuse crisis), integrated behavioral health support, peer support program referrals, community resource connections, follow-through tracking on psychosocial referrals (was the patient seen, what was the outcome).

============================================================ PHASE 7: INTERVENTION TRIGGER OPTIMIZATION ============================================================

Step 7.1 -- Trigger Threshold Calibration

Evaluate: how alert thresholds are set (fixed rules, data-driven optimization, clinician-configured), threshold sensitivity tuning (too sensitive = alert fatigue, too conservative = missed setbacks), threshold variation by risk level (higher-risk patients get more sensitive monitoring), threshold variation by recovery phase (early post-operative vs. late rehabilitation), threshold evaluation methodology (sensitivity and specificity at current thresholds).

Step 7.2 -- Intervention Menu

Evaluate: intervention options mapped to risk level (low risk: enhanced monitoring; moderate risk: treatment plan modification, increased frequency, additional modalities; high risk: physician notification, care conference, setting change consideration), intervention timing recommendations (how quickly should intervention occur after trigger), intervention escalation pathway, intervention documentation requirements.

Step 7.3 -- Intervention Effectiveness

Evaluate: whether interventions are tracked with outcomes (did the intervention prevent the setback), pre/post intervention metrics comparison, time from trigger to intervention measurement, intervention completion rate, clinician adoption of recommended interventions, false alarm rate (triggers where no setback would have occurred), cost-effectiveness of early intervention vs. setback management.

Write analysis to docs/setback-predictor-analysis.md (create docs/ if needed).

============================================================ SELF-HEALING VALIDATION (max 2 iterations) ============================================================

After producing output, validate data quality and completeness:

1. Verify all output sections have substantive content (not just headers).
2. Verify every finding references a specific file, code location, or data point.
3. Verify recommendations are actionable and evidence-based.
4. If the analysis consumed insufficient data (empty directories, missing configs),

note data gaps and attempt alternative discovery methods.

IF VALIDATION FAILS:

• Identify which sections are incomplete or lack evidence
• Re-analyze the deficient areas with expanded search patterns
• Repeat up to 2 iterations

IF STILL INCOMPLETE after 2 iterations:

• Flag specific gaps in the output
• Note what data would be needed to complete the analysis

============================================================ OUTPUT ============================================================

Setback Predictor Analysis Complete

• Report: docs/setback-predictor-analysis.md
• Risk factor model components evaluated: [count]
• Early warning indicators assessed: [count]
• Readmission prediction capabilities: [count]
• Adherence correlation metrics reviewed: [count]
• Psychosocial factors integrated: [count]
• Intervention trigger mechanisms analyzed: [count]

Critical findings:

1. [finding] -- [patient safety impact]
2. [finding] -- [prediction accuracy concern]
3. [finding] -- [intervention timing gap]

Top recommendations:

1. [recommendation] -- [expected improvement in setback prevention]
2. [recommendation] -- [expected improvement in prediction accuracy]
3. [recommendation] -- [expected improvement in intervention effectiveness]

NEXT STEPS:

• "Run /recovery-metrics to evaluate the outcome measurements that feed setback prediction."
• "Run /therapy-personalization to analyze how setback predictions drive treatment modifications."
• "Run /healthcare-compliance to verify prediction model compliance with clinical standards."

DO NOT:

• Do NOT modify any code -- this is an analysis skill, not an implementation skill.
• Do NOT include real patient data, medical records, or protected health information in output.
• Do NOT evaluate the prediction model as a replacement for clinical judgment -- it is a decision support tool.
• Do NOT ignore model bias -- prediction models trained on biased historical data will underserve certain populations.
• Do NOT treat non-adherence as solely patient responsibility -- systemic barriers (transportation, cost, work schedule) drive much non-adherence.
• Do NOT overlook psychosocial factors -- depression and catastrophizing are among the strongest predictors of poor rehabilitation outcomes.
• Do NOT set intervention thresholds without considering alert fatigue -- overwhelmed clinicians ignore alerts.
• Do NOT evaluate readmission prediction without examining discharge planning -- many readmissions are preventable with better transitions.
• Do NOT assume wearable data is always reliable -- device wear compliance, battery life, and sensor accuracy vary widely.

============================================================ SELF-EVOLUTION TELEMETRY ============================================================

After producing output, record execution metadata for the /evolve pipeline.

Check if a project memory directory exists:

• Look for the project path in ~/.claude/projects/
• If found, append to skill-telemetry.md in that memory directory

Entry format:

### /setback-predictor — {{YYYY-MM-DD}}
- Outcome: {{SUCCESS | PARTIAL | FAILED}}
- Self-healed: {{yes — what was healed | no}}
- Iterations used: {{N}} / {{N max}}
- Bottleneck: {{phase that struggled or "none"}}
- Suggestion: {{one-line improvement idea for /evolve, or "none"}}

Only log if the memory directory exists. Skip silently if not found. Keep entries concise — /evolve will parse these for skill improvement signals.