analysis-methods

von nvidia

Bringt dem Analysten-Agenten bei, korrekten und robusten Python-Analysecode für FHIR-klinische Daten mit pandas, matplotlib und scipy zu schreiben.

npx skills add https://github.com/nvidia/dgx-spark-playbooks --skill analysis-methods

Analysis Code Guidelines

FHIR Helpers Library

Always import the helpers library at the top of every analysis script:

import sys
sys.path.insert(0, '/sandbox/clinical-intelligence/skills/analysis-methods/scripts')
from fhir_helpers import *

Available functions

FunctionUse forHTTP calls
get_patients_with_condition(snomed_code)Find patients with a condition → list of IDs1-2
get_latest_labs_batch(loinc_code, patient_ids)Labs for a cohort → dict: pid → (value, unit, date)1-2
get_all_medications_batch(patient_ids)Meds for a cohort → dict: pid → [med names]1-2
build_cohort_df(patient_ids, loinc, lab_name, drug_check_fn)Full DataFrame with labs + meds2-3
get_latest_lab(patient_id, loinc_code)Lab for ONE patient → (value, unit, date)1
get_medications(patient_id)Meds for ONE patient → [names]1
get_latest_bp(patient_id)BP for ONE patient → (sys, dia, date)1-2
check_drug_class(med_list, drug_names)Check if any med matches drug list → bool0
fhir_get(path, params)Raw FHIR GET → parsed JSON1
get_all_pages(path, params)Paginated FHIR GET → all entries1+
save_chart_to_canvas(fig, filename)Save matplotlib figure to canvas directory0

Performance rules

  • Cohort queries (2+ patients): Use get_latest_labs_batch() and get_all_medications_batch(). These make 1-2 HTTP calls total regardless of patient count.
  • Single patient: Use get_latest_lab(), get_medications(), get_latest_bp().
  • NEVER loop over patients calling get_latest_lab() per patient. Each HTTP call through the sandbox proxy adds 1-3s. For 48 patients = 48 calls = 2+ minutes. The batch function does it in one call.

Execution Rules

  • Run scripts with python (NOT python3)
  • Write a SINGLE Python script for the entire task
  • Write the script to /tmp/<name>.py, then execute it
  • All HTTP inside the sandbox must use subprocess.run(["curl", ...]) — the requests library does NOT work

Mandatory Workflow

STEP 1 - WRITE SCRIPT (import fhir_helpers, write analysis)
STEP 2 - VALIDATE: python /sandbox/clinical-intelligence/scripts/validate_and_run.py --validate-only /tmp/<name>.py
STEP 3 - EXECUTE: python /tmp/<name>.py
STEP 4 - INTERPRET: explain results using clinical-knowledge skill

Code Structure

  1. Imports (always start with fhir_helpers import)
  2. Data collection (use batch functions)
  3. DataFrame construction
  4. Analysis (filters, aggregations)
  5. Visualization -- use save_chart_to_canvas(fig, filename) (NOT plt.savefig)
  6. Summary (print findings)
  7. Disclaimer

Care Gap Analysis Pattern

# Example: diabetes care gap
patients = get_patients_with_condition("44054006")  # SNOMED for diabetes
df = build_cohort_df(patients, "4548-4", "HbA1c",
                     lambda meds: check_drug_class(meds, ["metformin", "insulin", "glipizide"]))

gap = df[(df['HbA1c'] > 9) & (~df['on_target_med'])]
denom = len(df[df['HbA1c'].notna()])
pct = f"{len(gap)/denom*100:.1f}%" if denom > 0 else "N/A (no HbA1c data)"
print(f"Care gap: {len(gap)}/{denom} ({pct})")

Visualization

Always use dark theme. Use save_chart_to_canvas() instead of plt.savefig() directly.

import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt

plt.style.use('dark_background')
fig, ax = plt.subplots(figsize=(10, 6))
fig.patch.set_facecolor('#1a1a1a')
ax.set_facecolor('#1a1a1a')

# Histogram with NVIDIA green
ax.hist(values, bins=15, color='#76B900', edgecolor='#1a1a1a', alpha=0.85)
ax.axvline(x=threshold, color='#ff4444', linestyle='--', linewidth=2, label=f'Threshold ({threshold})')
ax.set_title("Title", fontsize=14, fontweight='bold', color='white')
ax.legend()
ax.grid(axis='y', alpha=0.2, color='#444444')
ax.text(0.98, 0.95, f"N = {len(values)}", transform=ax.transAxes, fontsize=11, color='#888888', ha='right', va='top')

# MANDATORY: use save_chart_to_canvas (NOT plt.savefig)
save_chart_to_canvas(fig, "chart.png")
plt.close()

Guardrails

  • Never compute statistics on fewer than 5 data points
  • Always report sample size: "45.0% (27 out of 60)"
  • Flag data quality issues if >30% missing
  • Do not fabricate data — report what exists, flag what's missing
  • All charts must include N annotation

Output Format

End every script with:

print(f"\nDisclaimer: This analysis is for research and operational purposes.")
print("Clinical decisions should be made by qualified clinicians.")

Mehr Skills von nvidia

compileiq-debug
nvidia
Verwenden, wenn etwas nicht stimmt: Search() hängt, alle Evaluierungen geben INVALID_SCORE zurück, Scores verbessern sich nicht, jede Konfiguration liefert dieselbe Zahl, ptxas-Fehler…
create-github-pr
nvidia
Erstelle GitHub-Pull-Requests mit der gh CLI. Verwende, wenn der Benutzer einen neuen PR erstellen, Code zur Überprüfung einreichen oder einen Pull-Request öffnen möchte. Auslöser-Schlüsselwörter -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Scannt andere offene Issues, um solche zu finden, die ein bestimmter PR möglicherweise ebenfalls behebt oder versehentlich kaputt macht. Gibt benachbarte Fix-Möglichkeiten und Widerspruchsrisiken mit Datei:Zeile… aus.
fhir-basics
nvidia
Bringt Agenten bei, wie FHIR R4 APIs funktionieren, welche Ressourcen verfügbar sind, wie man sie mit Suchparametern abfragt und wie man alle Antwortformate korrekt parst…
compileiq-validate-result
nvidia
Verwende NACH Abschluss einer Suche und VOR dem Einfordern eines Speedups oder dem Versand eines ACF. Lädt die dump_results CSV, extrahiert Top-K-Kandidaten (Einzelziel)...
changelog-audit
nvidia
Auditiere die CHANGELOG.md vor einem Release: stelle verlorene Einträge wieder her, sortiere nach Benutzerauswirkung, verfeinere die Sprache der Einträge, führe Zeilenumbrüche durch und (im Release-Branch-Modus) erhöhe die Vergleichsnummer…
maintain-dynamic-plugins
nvidia
Verwalte NeMo Relay dynamische Plugin-Lader, Manifeste, Rust native SDKs, gRPC Worker-Protokoll, Python Worker-SDK, Dokumentation, Tests und Abdeckung des Release-Workflows
dgx-diagnose
nvidia
Diagnostizieren Sie häufige DGX Station GB300-Probleme – CUDA-Abstürze, falsche GPU-Zuweisung, vLLM/SGLang-Container-Fehler, MIG-Status-Probleme, NVLink/Fabric-Manager-Fehler,…