Assigning International Classification of Diseases (ICD-9) codes is a critical task in modern healthcare, underpinning billing, clinical research, and the development of treatment guidelines. However, for medical professionals in Intensive Care Units (ICUs), manual assignment is a tedious and time-consuming process. ICU data is high-dimensional, and patients often present with multiple, overlapping conditions, making automated assignment a significant challenge.
Existing automated systems often struggle with three primary issues: data redundancy in medical charts, the assumption of feature uniformity (treating all symptoms as equally relevant to every disease), and label independence (ignoring the natural co-occurrence of certain conditions). To address these challenges, the Integrated Framework of Feature Learning and Classification (IFFLC) was developed to automate the assignment of multiple diagnosis codes with high accuracy and interpretability.
The IFFLC Framework
The IFFLC framework adopts a two-stage approach to transform raw clinical data into precise diagnostic codes.
1. Feature Selection via Disease Information Entropy
To manage the high dimensionality of ICU data—which can include over 2,000 numerical features from charts and labs—the framework employs Fuzzy Mutual Information (FMI). This entropy-based method identifies a subset of features that maximizes relevance to the diagnosis labels while minimizing redundancy. The study found that using just 200 selected features provided an optimal balance between computational efficiency and classification performance.
2. The IFFLC Classifier
The core innovation of the framework is a multi-label classifier designed as an optimization problem with two key regularizers that mimic "doctor's intuition":
- Sparsity ($l_1$-norm): This forces the model to identify "disease-specific features," ensuring that only the most discriminative markers for a particular disease are utilized.
- Correlation Embedding: The model incorporates a correlation matrix to capture how diseases relate to one another, allowing the system to "share" information between frequently comorbid conditions, such as hypertension and heart disease.
The resulting optimization problem is solved using the Alternating Direction Method of Multipliers (ADMM), ensuring efficiency even with large-scale clinical datasets.
Experimental Results and Insights
The framework was validated using the MIMIC-III database, focusing on 5,058 patient records involving diseases of the circulatory system (ICD-9 codes 390-459). IFFLC was compared against six state-of-the-art methods, including Binary Relevance (BR) and ML-KNN.
- Superior Accuracy: IFFLC achieved an Average Precision of 0.7521, significantly outperforming the next best competitor, LIFT (0.7252).
- Lower Error Rates: The model achieved the lowest Hamming Loss (0.0288) and Ranking Loss (0.0432) among all tested methods.
- Stability: Unlike many models that excel in one metric but fail in others, IFFLC showed remarkable stability across six different multi-label evaluation criteria.
Conclusion and Future Work
By integrating feature learning with disease correlation, IFFLC provides a robust and interpretable solution for automated medical coding. While the current model focuses on second-order (pairwise) correlations, future work aims to model higher-order relations and complex hierarchies within the ICD coding structure to further improve diagnostic precision across a broader range of medical categories.
Summary updated