This paper, presented at the NeurIPS 2018 Workshop on Machine Learning for Health (ML4H), explores a pragmatic approach to predicting the progression of diabetes by leveraging a data source often overlooked in clinical research: financial billing records.
TL;DR
Researchers developed a self-attentive Recurrent Neural Network (RNN) that predicts severe diabetes complications—such as amputations and kidney failure—up to eight months in advance. By using standardized financial claims data instead of messy clinical notes, the model provides a scalable way for healthcare providers to identify high-risk patients and intervene before a crisis occurs.
What Problem Does This Paper Address?
Diabetes is a global epidemic, affecting over 420 million people. While the disease is manageable, its progression can lead to devastating complications like heart disease, strokes, and organ failure. To prevent these, healthcare providers need "early warning systems."
Most existing predictive models rely on Electronic Health Records (EHRs), which contain detailed clinical data like BMI, blood pressure, and lab results. However, in many parts of the world—particularly in developing nations—EHR data is fragmented, non-digital, or too heterogeneous to integrate. Financial records (medical claims), on the other hand, are highly standardized for billing purposes and are often centralized by insurance providers. The challenge is that financial data is "noisier" and sparser than clinical data; it tells you that a patient was billed for a needle, but not necessarily what their blood sugar level was.
Methods
The authors treated a patient’s history of medical billing codes as a sequence, much like words in a sentence.
- Data Encoding: They used the TUSS (Supplementary Healthcare Unified Terminology) standard, a Brazilian national standard for medical procedures.
- Embeddings: To handle the 150,000+ unique codes, they used Word2Vec skip-grams. This allowed the model to learn "semantic" relationships between codes—for example, recognizing that different doses of the same drug are related, or that certain diagnostic tests often precede specific surgeries.
- Architecture: The core model is a bidirectional Long Short-Term Memory (LSTM) network equipped with a self-attention mechanism.
- The LSTM captures the temporal evolution of the disease.
- The self-attention layer allows the model to "ignore" irrelevant billing noise (like basic hospital supplies) and focus on the specific sequence of events that signal a high risk of complication.
Datasets
The study utilized a massive dataset from a Brazilian health plan provider:
- Scale: Five years of data covering approximately 7 million unique individuals and 327 million records.
- Filtering: Because the dataset lacked specific diagnosis codes (ICD-10) for all patients, the researchers identified a cohort of 100,000 "likely diabetics" based on those who had received at least two glycated hemoglobin (HbA1c) tests within a year.
- Target: The model was trained to predict three specific acute complications: (i) Amputations/debridements, (ii) Revascularization/angioplasty, and (iii) Hemodialysis.
Key Insights and Ideas
- The "Language" of Claims: The most striking idea is treating medical billing as a natural language processing (NLP) problem. By pre-training embeddings on codes, the model learns the "grammar" of disease progression without needing a doctor to manually define features.
- Attention Filters the Noise: Financial data is cluttered with administrative entries. The self-attention mechanism proved vital because it can assign higher weights to a specific cardiovascular exam while down-weighting the "noise" of routine office supplies billed at the same time.
- Standardization as a Bridge: Because financial codes are often mapped to international standards (like CPT or LOINC), a model trained on one country's billing data could theoretically be adapted to another with minimal friction compared to clinical notes.
Key Results
The researchers tested the model’s ability to predict complications across four different "time gaps" (60, 120, 180, and 240 days).
- Performance: The self-attentive model (LSTM+SA) significantly outperformed a standard LSTM baseline.
- AUC Scores: The initial model achieved an Area Under the ROC Curve (AUC) between 0.81 and 0.83.
- Advanced Results: In a "work-in-progress" update mentioned in the paper, the authors noted that by using a novel method to embed dates, they improved the model's performance to an AUC of 0.94.
- Lead Time: The model maintained high accuracy even when predicting complications 240 days (8 months) into the future, though performance naturally dipped slightly as the prediction window lengthened.
Limitations and Future Work
- The "Silent" Patient: The authors acknowledge that if a patient’s disease is progressing but they aren't seeking treatment (and thus generating no bills), the model has no data to work with.
- False Positives: While the AUC is high, the model still produces false positives. In a real-world setting, "alarm fatigue" for healthcare providers is a concern if the system flags too many patients who aren't actually at immediate risk.
- Clinical Nuance: Financial records lack the "why" behind a procedure. Future work could involve hybrid models that combine these broad financial snapshots with deep clinical data where available.
- Deployment: The ultimate goal is to integrate this into a monitoring program where insurers can proactively offer personalized prevention plans to high-risk individuals.
Summary updated