EarlyAlertAI: Patient Deterioration Detection

Inspiration

Unexpected patient deterioration in ICUs is one of the most critical challenges in healthcare. Many patients show subtle warning signs hours before clinical collapse, but these signals are often missed due to workload, fragmented monitoring, and alert fatigue. We were inspired to build a decision-support tool that helps clinicians detect early risk patterns using routine ICU data.

What it does

EarlyAlertAI is an interpretable early-warning system that predicts the risk of patient deterioration hours in advance. It provides:

->Risk classification (Low / Medium / High)

->Early detection of critical cases with high recall

->Explainable alerts showing the main contributing vital trends

->Reduced unnecessary alarms to prevent alert fatigue

This tool is designed to support clinicians, not replace medical judgment.

How we built it

We developed the project using publicly available de-identified ICU time-series data (eICU demo dataset).

Steps included:

->Merging patient demographics with periodic and aperiodic vital measurements

->Creating clinically meaningful deterioration labels based on worsening physiological thresholds

->Performing EDA to uncover early patterns in heart rate, blood pressure, and oxygen saturation

->Training and evaluating multiple ML models

->Finalizing an interpretable pruned Decision Tree classifier with strong recall performance

->Adding risk tiers and feature-based explanations for doctor trust

Challenges we ran into

->ICU data is noisy, incomplete, and irregularly sampled

->Defining deterioration labels without direct outcome annotations required careful clinical reasoning

->Balancing patient safety (high recall) with reducing false alarms (precision)

->Handling missing values realistically to match real hospital conditions

Accomplishments that we're proud of

->Built a complete ICU deterioration prediction pipeline within hackathon constraints

->Achieved high test recall while minimizing missed critical cases

->Reduced alert fatigue through balanced classification thresholds

->Delivered an interpretable decision-support model suitable for clinical workflows

What we learned

->Healthcare AI must prioritize safety, interpretability, and responsible deployment

->Real-world hospital data is far more complex than clean benchmark datasets

->Decision-support framing is essential for trust, ethics, and adoption

->Even simple models can provide strong value when paired with meaningful clinical features

What's next for EarlyAlertAI: Patient Deterioration Detection

Future improvements include:

->Validation on full credentialized ICU datasets such as MIMIC-IV

->Incorporating lab trends and richer temporal modeling (LSTMs/Transformers)

->Building a clinician-facing dashboard for real-time monitoring

->Conducting structured feedback sessions with ICU professionals

->Strengthening GDPR-compliant deployment and hospital integration pathways

Built With

• collaborative
• database
• demo)
• eicu
• icudataset)
• machine-learning
• matplot
• numpy
• pandas
• physionet
• python
• research
• scikit-learn
• seaborn