convez.ai
PluginDados & AnalyticsIntermediário
muhammed-healthai

AuditAI — Assistente de Auditoria Clínica com IA

muhammed-healthai/Audit-AI

Transforma dados de auditoria clínica em relatórios com recomendações IA fundamentadas

Ver no GitHub
InicialPluginScore · de 0 a 100

Ferramenta que converte dados de auditoria hospitalar (CSV/Excel) em relatórios PDF no padrão NHS com recomendações personalizadas geradas por Claude, fundamentadas em diretrizes NICE/RCP/WHO. Detecta automaticamente o tópico da auditoria e gera sugestões específicas em ~10 segundos.

Instalação

$npx convez-ai muhammed-healthai/Audit-AI

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 6 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Gerar recomendações personalizadas para auditorias de conformidade clínica
  • Automatizar produção de relatórios PDF padronizados para hospitais NHS
  • Analisar gaps de compliance em protocolos de VTE, higiene de mãos, prescrição de medicamentos

Veredito

Plugin técnico robusto que demonstra uso legítimo de LLM em contexto clínico — recomendações realmente contextualizadas por auditoria, não boilerplate. Pronto para uso por hospitais e residentes médicos; excelente exemplar de integração Claude + Streamlit com valor real.

README

A tradução para o português deste README ainda está na fila.

AuditAI — AI-Powered Clinical Audit Assistant

Live demo Python 3.11 Built with Claude License: MIT

A Streamlit app that turns a ward-level audit CSV into an NHS-styled PDF report — with topic-aware, NICE/RCP-grounded recommendations generated by Claude, not hardcoded boilerplate.

🔗 Live demo: audit-ai-mo.streamlit.app

Built by a resident doctor who got tired of writing the same audit recommendations by hand.


What it does

Upload a CSV / Excel of audit data → AuditAI:

  1. Auto-detects the boolean compliance fields (Yes/No, True/False, 1/0)
  2. Computes overall and per-component compliance with visual bars
  3. Detects the audit topic from the column names (e.g. "VTE risk assessment", "Hand hygiene (WHO 5 Moments)", "Inpatient medication prescribing")
  4. Generates 3-4 tailored recommendations via Claude, citing the relevant UK guidance (NICE NG89, WHO 5 Moments, Royal Pharmaceutical Society, etc.)
  5. Exports an NHS-styled PDF with summary, colour-coded compliance bars, AI recommendations, and a sign-off section

The whole flow takes ~10 seconds end-to-end.


The "AI" earns its name

Most "AI"-branded portfolio projects use hardcoded if/else recommendations dressed up in marketing language. AuditAI's recommendations are generated per-audit by Claude Sonnet 4.6, with the audit topic detected from column names.

What the same compliance gap produces

For a VTE audit (vte_done, prophylaxis_given, compliant):

"Configure the EPR admission clerking template to include a mandatory VTE risk assessment field using the NICE-recommended tool (NICE NG89), preventing form submission until the assessment is completed — this directly targets the 30% gap in vte_done compliance."

For a hand hygiene audit (before_patient_contact, after_patient_surroundings, etc.):

"Launch a targeted education session specifically addressing Moments 1 and 5 (before patient contact and after patient surroundings), which scored lowest at 65% and 60% respectively, using WHO 5 Moments for Hand Hygiene training materials and NICE guideline PH69 as the framework."

For a drug chart audit (prescriber_signature, weight_recorded, prescriber_gmc_number, etc.):

"Embed a mandatory weight-recording field in the EPR admission clerking proforma, so prescriptions cannot be finalised without a documented weight — addressing the 33.3% gap in weight_recorded, in line with NICE medicines optimisation guidance (NG5)."

Same app. Different audits. Different guideline bodies. Different specific gaps targeted. That's what real LLM use looks like in a clinical tool.


Try it with example data

Three example datasets ship in examples/ so you can see the topic detection switch live:

FileAudit topicWhat Claude detects
vte_prophylaxis_audit.csvVTE prophylaxis prescribingCites NICE NG89, NHS CQUIN VTE framework
hand_hygiene_audit.csvHand hygiene complianceCites WHO 5 Moments, NICE PH69, NICE NG125
drug_chart_audit.csvInpatient prescribingCites NICE NG5, RPS Prescribing Competency Framework

Upload any one to the live demo and see how the recommendations adapt.


Generated PDF report

Each report includes the auto-detected topic, colour-coded compliance bars (red <75%, amber 75–95%, green ≥95%), Claude-generated recommendations with their guideline framework header, and a clinician sign-off section.


Architecture

Audit-AI/
├── app/
│   ├── streamlit_app.py    # UI + boolean detection + compliance maths
│   ├── llm.py              # Claude topic detection + recommendation prompt
│   └── report_utils.py     # ReportLab PDF generation + compliance bars
├── examples/               # 3 example audit datasets
├── assets/                 # NHS logo + README screenshots
├── requirements.txt
└── README.md

The LLM module is decoupled — it loads its API key from .env (local) or st.secrets (Streamlit Cloud), and falls back gracefully to generic recommendations if the API is unavailable. The Streamlit UI shows a transparent badge (✨ Generated by Claude / ⚠️ Fallback) so users always know whether they're seeing real AI output.


Run locally

git clone https://github.com/M-Omarjee/Audit-AI.git
cd Audit-AI

python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# Add your Anthropic API key (get one at console.anthropic.com)
echo 'ANTHROPIC_API_KEY=sk-ant-...' > .env

streamlit run app/streamlit_app.py

Then upload examples/vte_prophylaxis_audit.csv to see it work.


Limitations

  • Single time-point audits only — doesn't yet support comparing audit cycles (cycle 1 vs cycle 2 trend graphs are on the roadmap)
  • Synthetic example data — the shipped examples are illustrative; real audit data should be anonymised before upload (the deployed app does not log uploads, but defence-in-depth)
  • LLM output is generated, not curated — recommendations are clinically literate but should be reviewed by the user before submission. Claude can be wrong.
  • No persistence — each audit is processed fresh; no user accounts, no saved history. By design for a portfolio MVP.

Roadmap

  • Multi-cycle audit comparison (cycle 1 → intervention → cycle 2 trend)
  • Free-text "Notes" column summarisation (extract themes from clinician comments)
  • Export-to-DOCX option alongside PDF
  • Custom NHS Trust branding (logo + colour overrides)
  • Optional NICE guideline link-out in PDF recommendations

Author

Dr Muhammed Omarjee Resident Doctor (MBBS, King's College London 2023) Building practical AI tools for NHS frontline workflows.

Sister projects:

  • ECG-Explain — 12-lead ECG classifier with Grad-CAM interpretability (PTB-XL, AUROC 0.91)
  • Sepsis-AI — NEWS2 vs ML benchmark with decision curve analysis and subgroup audit

License

MIT

Ver o original no GitHub

Bom para

Relacionados