Medical database extraction and analysis

Inspiration

Clinical microbiology reports and antimicrobial-resistance databases contain valuable information, but they are often fragmented across text reports, CSV files, public surveillance databases, and model artifacts. We wanted to build a reproducible tool that helps researchers and clinical microbiology teams extract, standardize, analyze, and explain Acinetobacter baumannii antimicrobial resistance data in one workflow.

What it does

This project provides a GitHub-testable AMR prediction and report interpretation system for Acinetobacter baumannii. It can predict resistance probabilities, add Bio-ab contextual interpretation, score newer agents with transparent exploratory rules, rank candidate single-agent and combination-therapy options for research review, monitor public AMR databases, extract AST information from microbiology report text, infer likely Bio-ab genotype patterns, and generate CSV, Markdown, Chinese Word, and English Word reports.

How we built it

The system is implemented as a Python command-line toolkit with reproducible sample inputs, reference tables, pretrained XGBoost/joblib model artifacts, pytest coverage, and GitHub Actions CI. The workflow is modular: prediction, database monitoring, report interpretation, candidate ranking, and Word report generation can be tested independently.

Repository: https://github.com/huyongyu123456789/Acinetobacter-baumannii-prediction-and-drug-sensitivity-interpretation

Challenges we ran into

The main challenge was making the system useful while keeping its boundaries safe. Public AMR data can contain duplicate isolates, incomplete metadata, inconsistent AST interpretation standards, MIC censoring, geographic sampling bias, and annotation-version differences. For that reason, database monitoring does not automatically retrain or replace models. It produces review queues and reports so model updates remain human-reviewed.

Another challenge was report output. A mixed bilingual report is useful for some teams, but clinical research workflows often need separate Chinese and English documents. Version 1.5.1 solves this by adding split Word reports while preserving Markdown and CSV outputs.

Accomplishments that we're proud of

• A complete AMR prediction and interpretation workflow is available as a public GitHub project.
• The project includes CI-ready tests and reproducible sample files.
• Report interpretation now produces independent Chinese and English Word reports.
• The tool keeps clinical-safety boundaries explicit: candidate therapies and concentrations are research-only and are not clinical prescribing recommendations.

What we learned

We learned that medical database extraction is not only a modeling problem. The surrounding workflow matters just as much: data cleaning, schema design, interpretability, audit trails, human review, output formats, and safety language all affect whether a research tool can be responsibly used and shared.

What's next

Next steps include stronger OCR support for scanned reports, richer database connectors, external validation datasets, calibrated model updates, improved phenotype-standardization workflows, and a future Version 2.0 model release after full retraining and external validation.

Safety note

This project is for research and decision-support exploration only. It does not replace phenotypic antimicrobial susceptibility testing, MIC confirmation, PK/PD assessment, hospital guidelines, infectious disease consultation, clinical pharmacy review, or clinician judgment.

Built With

• amr
• bioinformatics
• clinical-microbiology
• csv
• github-actions
• health
• joblib
• markdown
• numpy
• pandas
• public
• pytest
• python
• python-docx
• scikit-learn
• xgboost