Predicting Sepsis with Analytics - Myridius and Boston College Collaboration
Myridius and Boston College (BC) recently hosted an analytics competition, challenging BC students to predict sepsis using panel data. The competition featured a large dataset: 1 million observations from 28,000 patients for training, and 454,000 observations from 6,000 patients for testing. With only 2% of cases labeled sepsis-positive, participants tackled severe class imbalance.
The Challenge and Evaluation
The primary success metric was the F1-score, which balances precision and recall. Judges also assessed the overall presentation, innovation in modeling, clarity of communication, and understanding of the problem through a detailed evaluation rubric.
Winning Solution
The top team’s solution, based on CRISP-DM methodology, excelled in understanding and innovation. Key steps included:
- Preprocessing: Data cleaning, imputation, feature engineering, train-validation splits, and oversampling.
- Model Testing: Evaluating logistic regression, tree models, neural networks, RNNs, LSTMs, and GRUs. An RNN achieved the best F1-score.
- Analysis: Thoughtful predictions and insights.
Despite not achieving the highest F1-score (“0.21”), their approach, communication, and insights set them apart.
Results and Impact
Myridius awarded $3,000 in prizes to the top teams, providing students with hands-on experience in tackling healthcare challenges. The competition reinforced the importance of analytics in transforming healthcare outcomes, aligning with Myridius' mission to foster data-driven innovation.