Introduction
Predicting mortality risk for patients with Acute Myocardial Infarction (AMI)—commonly known as a heart attack—is critical for providing tailored, life-saving care. Traditional scoring systems like APACHE often rely on a limited number of variables and require frequent manual recalibration. While Machine Learning (ML) has been applied to Electronic Health Records (EHRs) to address these gaps, most existing models focus on short-term survival (e.g., 24 hours or 30 days) and overlook "unstructured" data. This study demonstrates that combining structured clinical data with the rich narrative found in discharge summaries significantly improves one-year mortality predictions, achieving an accuracy of 92.89%.
Methods and Data Integration
The researchers developed a "mixed data" approach that fuses traditional structured features with text-based embeddings extracted from clinical notes.
Feature Extraction and Modeling
- Structured Data: 79 features were utilized, including demographics, lab results, and comorbidities.
- Unstructured Data: Discharge summaries were transformed into dense vector representations using Word2Vec (Skip-gram model). These embeddings were averaged to create a single "document embedding" for each patient admission.
- The Model: A fully connected deep neural network was implemented with a 279-dimension input layer, two hidden layers (400 neurons each), and a softmax classification layer. To prevent overfitting, the architecture incorporated L2 regularization, Dropout (0.3), and batch normalization.
Dataset
The study utilized the MIMIC-III database, identifying a cohort of 5,436 hospital admissions related to AMI. The dataset was imbalanced, with approximately 30% of patients having died within one year post-admission.
Key Results
The transition from structured-only data to "mixed" data (structured + text) yielded impressive gains, particularly for the Deep Learning (DL) architecture. While traditional "shallow" learners like Logistic Model Trees (LMT) showed marginal improvement when text was added, the DL model's performance surged.
As shown in the results, the DL model achieved a 92.89% accuracy and a 0.928 F-measure. Notably, recall improved significantly, meaning the model became much better at correctly identifying high-risk patients, thereby reducing dangerous false negatives.
Key Insights and Discussion
- Deep Learning Synergy: The results suggest that DL is uniquely capable of extracting complex, non-linear patterns from high-dimensional text data that shallow models miss.
- Feature Synergy: Combining text with admission and demographic information (age, gender, initial ER diagnosis) provided a stronger predictive boost than combining text with lab results or treatments alone.
- Long-Term Signal: This research proves that data collected during an ICU stay contains a "signal" strong enough to predict a patient's health status 365 days into the future, moving beyond immediate crisis management.
Limitations and Future Work
The authors noted that using ICD-9 codes can introduce noise due to varying coding practices. Future work will explore Recurrent Neural Networks (RNNs) or LSTMs to capture the temporal sequence of clinical notes, which is currently lost by averaging embeddings. Additionally, the team aims to utilize the Unified Medical Language System (UMLS) to extract specific medical concepts for even cleaner predictive signals.
Summary updated