Can routinely available ICU information help identify high-risk sepsis patients?
The study uses first-24-hour clinical features and supervised machine-learning baselines to evaluate mortality prediction using hospital_expire_flag as the outcome.
Healthcare AI
Clinical AI research grounded in real data and explicit evidence limits.
My healthcare work uses real-world clinical datasets to study prediction, NLP, cohort construction, model evaluation, and the steps required to move from retrospective analysis toward trustworthy clinical research.
14,966Adult sepsis patients in the MIMIC-III cohort
15Structured first-24-hour clinical features
3Primary supervised model families evaluated
Real dataPhysioNet MIMIC-III via BigQuery; no synthetic demo cohort
Flagship study: sepsis mortality risk
A reproducible workflow centered on adult sepsis patients and hospital mortality. The objective is methodological learning and evidence generation—not a claim of clinical readiness.
Real MIMIC-III critical-care records
The cohort includes 14,966 adult sepsis patients. Feature groups include vitals, laboratory values, GCS, SOFA, age, gender, and comorbidity-related information.
Baseline-to-nonlinear comparison
Logistic Regression, Random Forest, and XGBoost provide interpretable and nonlinear comparison points. The workflow includes imputation, scaling where appropriate, train/test separation, class-imbalance handling, ROC/AUC, and confusion-matrix evaluation.
Retrospective research only
No result on this site should be interpreted as a clinically validated medical device, treatment recommendation, or prospective patient-care system.
Clinical research pipeline
From EHR data to evidence
MIMIC-III / EHR data
→
Cohort definition
→
Structured features + notes
→
Models / NLP
→
Validation
→
Calibration / bias / limits
→
Research evidence
The deployment boundary comes after validation and governance—not before.
Clinical NLP work
A separate research workflow uses real MIMIC-III NOTEEVENTS joined with sepsis-related diagnoses. The goal is to compare clinical NLP tooling and representation methods on authentic notes.
General NLP baseline
Entity extraction and linguistic processing provide a general-purpose benchmark.
Biomedical NLP
Scientific and biomedical language models support domain-relevant entity extraction and representation.
Clinical NLP
Clinical-oriented processing supports comparison with more general and biomedical pipelines.
Hospital collaboration ladder
Clinical collaboration should increase in sensitivity only as governance and evidence mature.
1. Public/de-identified research
Use PhysioNet and other approved de-identified datasets to establish methods, reproducibility, and preliminary evidence.
2. Hospital-defined research problem
Work with clinical stakeholders to define a useful question, success criteria, workflow constraints, and governance requirements.
3. Controlled institutional data access
Only after appropriate agreements, security controls, data-use terms, and IRB/privacy determinations are in place.
4. Prospective evaluation
Test whether the system works in the intended environment before making deployment or clinical-impact claims.
Privacy boundary: this public portfolio contains no PHI and should never be used to publish patient-level information, restricted MIMIC content, credentials, or institution-confidential data.