TL;DR
This paper investigates how demographic factors influence the performance of medical code prediction models, finding that models perform worse on groups with label distributions that significantly differ from the global data. They propose a novel "label distance metric" to quantify this difference, offering a new way to assess and potentially mitigate fairness issues in clinical NLP.
What Problem Does This Paper Address?
Automated medical code prediction, crucial for documentation and billing, faces a significant challenge: data imbalance. This means certain medical codes appear far more frequently than others in training datasets. While this imbalance is known to affect model performance across different codes, this paper explores a less-studied area: how demographic imbalances (age, gender, ethnicity, insurance type) in the data might lead to performance disparities across different patient groups. The core problem is that if a model is trained on data that doesn't accurately represent certain demographic groups, it might provide lower quality predictions for those groups, potentially leading to unequal healthcare outcomes.
Methods
The researchers employed a multi-faceted approach:
- Data Analysis: They first analyzed the MIMIC-III dataset to understand the distribution of demographic variables and their relationship with medical codes.
- Model Reproduction: To ensure a fair comparison, they reproduced a state-of-the-art medical code prediction model (Li and Yu, 2020), which uses a multi-filter convolutional neural network with residual connections and a label attention layer. They re-trained this model using specific early-stopping criteria.
- Novel Metric Development: A key contribution is the proposal of a "label distribution distance metric." This metric quantifies how much a specific demographic group's distribution of medical codes deviates from the overall dataset's distribution. A larger distance suggests a greater difference.
- Evaluation Metrics: Beyond standard micro- and macro-averaged F1 scores, they introduced sample-averaged F1 scores. This metric evaluates performance at the individual patient level, which they argue is more relevant for assessing fairness in healthcare. They also used error metrics like False Negative Rate (FNR) and False Positive Rate (FPR).
- Statistical Analysis: Kruskal-Wallis tests were used to compare performance across groups, and Pearson correlation was employed to assess the relationship between training data size, label distance, and model performance.
Datasets
The study utilized the MIMIC-III v1.4 dataset. Specifically, they focused on the MIMIC-III 50 dataset, which includes the top 50 most frequent ICD-9 codes. This subset was chosen for simplification of analysis. The dataset was split into training, validation, and testing sets following established protocols. Demographic information (age, gender, ethnicity, insurance type) was extracted from the dataset.
Key Insights and Ideas
- Demographic Imbalance is Real: The MIMIC-III dataset, like many real-world datasets, exhibits imbalances across demographic groups, particularly in age and insurance type.
- Performance Disparities Exist: The medical code prediction model showed noticeable performance differences across various demographic groups, especially for age and insurance type.
- Label Distribution is Key: The most striking finding is the strong negative correlation between a group's "label distance" (how much its ICD code distribution differs from the global distribution) and the model's performance on that group. This suggests that models struggle when a group's coding patterns are unusual compared to the majority.
- Training Data Size Matters Less Than Distribution: Counterintuitively, the correlation between the amount of training data for a group and its performance was weak and not statistically significant. This highlights that the nature of the data distribution within a group is more critical than just its sheer volume.
- Sample-Averaged Metrics are Crucial for Fairness: The authors advocate for sample-averaged metrics, arguing they better reflect individual patient outcomes than traditional label-averaged metrics, which can mask disparities.
Key Results
- Age and Insurance Groups Show Significant Differences: Kruskal-Wallis tests revealed statistically significant differences in sample-averaged F1 scores for age groups (p<0.001) and insurance types (p<0.001).
- Younger and Older Patients Perform Worse: Patients younger than 30 and older than 90 showed worse predictions.
- Specific Ethnicities and Insurance Types Lag: Hispanic and Asian patients, as well as those with Medicaid or "Other" insurance, received worse predictions compared to their counterparts.
- Strong Negative Correlation: A very strong negative correlation (-0.95, p<0.001) was found between label distance and performance. This means as a group's label distribution deviates more from the norm, its performance drops significantly.
- Reproduced Model Performance: The reproduced model achieved a macro-averaged F1 score of 59.2% and a sample-averaged F1 score of 60.6%, slightly underperforming the original model but within a reasonable range.
Limitations and Future Work
- Subset of Codes: The study focused on the top 50 ICD codes, which simplifies analysis but may not generalize to the full complexity of the MIMIC-III dataset (over 47,000 codes).
- Single Model: Only one state-of-the-art model was investigated. Future work should explore other models, especially those using pre-trained language models.
- Confounding Variables: Demographic variables are often correlated (e.g., age and insurance type), making it difficult to isolate the exact impact of each individual factor. Further causal analysis is needed.
- Addressing Label Shift: The authors suggest future research could explore techniques like adjusting training sample weights or adapting pre-trained classifiers to mitigate the observed label shift issues.
- Building Fair Models: A key future direction is to investigate how to build models that perform equally well across all demographic groups, regardless of their label distribution.
Summary updated