PheNorm: Unsupervised Gold-Standard Label Free Phenotyping Algorithm for EHR Data

Yu Sheng, Víctor M. Castro, Clara-Lea Bonzel, Molei Liu, Chuan Hong, Tianxi CaiOpenalex Unknown 2021 · DOI · OpenAlex

Abstract

The PheNorm algorithm combines predictive variables, such as counts of International Classification of Diseases (ICD) codes, with other Electronic Health Record (EHR) features like healthcare utilization and processed clinical note data to generate accurate risk prediction and disease classification scores. By normalizing surrogates to resemble a Gaussian mixture and leveraging auxiliary features through random corruption denoising, PheNorm provides a robust, unsupervised approach to phenotyping. This method eliminates the need for manually labeled "gold standard" datasets, addressing a significant bottleneck in medical informatics. Background and details about the method can be found at Yu et al. (2018) doi:10.1093/jamia/ocx111.

Introduction

In medical informatics, "phenotyping" is the critical process of identifying whether a patient has a specific condition based on their Electronic Health Record (EHR). Traditionally, building a phenotyping algorithm requires a "gold standard"—a set of patient records manually reviewed and labeled by clinicians. This manual review is a massive bottleneck: it is time-consuming, expensive, and difficult to scale across different hospitals or diseases. Researchers often face a "cold start" problem where they have massive amounts of data but no labeled examples to train a supervised model.

PheNorm addresses this by providing a label-free, unsupervised method to obtain accurate phenotyping scores, allowing researchers to classify patients using the data already present in the EHR.

Methodology

The PheNorm algorithm operates on several key technical pillars to transform raw EHR data into reliable phenotype probabilities:

  1. Surrogate Selection: The process begins with "surrogates"—highly predictive variables like the count of specific ICD codes or mentions of a disease in clinical notes (NLP data).
  2. Gaussian Mixture Normalization: Raw EHR counts are typically skewed. The algorithm normalizes these surrogates so they resemble a Gaussian mixture model. This effectively identifies two underlying distributions: one representing patients who likely have the condition and another for those who likely do not.
  3. Random Corruption Denoising: To enhance robustness, PheNorm incorporates additional EHR features such as healthcare utilization rates. It employs a "random corruption" technique—similar to the logic used in denoising autoencoders—to leverage these extra features without overfitting to the noise inherent in EHR data.
  4. Automated Scoring: The final output is a probability score for each patient, representing the likelihood of the phenotype, along with the beta coefficients used to reach that score.

Implementation and Parameters

The core functionality is implemented via the PheNorm.Prob function. This function requires specific inputs to manage the surrogates, utilization metrics, and the denoising process.

Parameters for the PheNorm.Prob function, defining surrogates, utilization metrics, and data requirements.

As shown in the parameter definitions, the algorithm requires all data columns to be log-transformed before processing. It also allows users to specify a corrupt.rate (defaulting to 0.3) for the denoising step and a train.size to manage the computational scale of the internal sampling.

Discussion and Limitations

PheNorm represents a significant shift toward automated EHR research by turning "noise into signal." By treating ICD codes as noisy surrogates rather than absolute truths, the algorithm uses statistical normalization to find the true signal underneath. The inclusion of healthcare utilization (e.g., note counts or encounter numbers) is a clever way to control for data density bias, where patients with more frequent interactions might otherwise appear to have higher disease counts simply due to more documentation.

However, users must be aware of certain limitations:

  • Surrogate Dependency: While unsupervised, the model's accuracy is still dependent on the quality of the user-provided surrogate. A poorly correlated surrogate will lead to inaccurate results.
  • Preprocessing Requirements: The necessity for log-transformation adds a mandatory preprocessing step for the user.
  • Computational Intensity: The default training size (10 times the number of rows) suggests that the algorithm performs significant internal simulation, which may require substantial memory for very large datasets.

Despite these constraints, PheNorm offers a powerful, scalable solution for researchers needing to phenotype large EHR datasets without the luxury of manual labels.

Summary updated

Abstract

The algorithm combines the most predictive variable, such as count of the main International Classification of Diseases (ICD) codes, and other Electronic Health Record (EHR) features (e.g. health utilization and processed clinical note data), to obtain a score for accurate risk prediction and disease classification. In particular, it normalizes the surrogate to resemble gaussian mixture and leverages the remaining features through random corruption denoising. Background and details about the method can be found at Yu et al. (2018) &lt;<a href="https://doi.org/10.1093%2Fjamia%2Focx111" target="_top">doi:10.1093/jamia/ocx111</a>&gt;.
BibTeX
@inproceedings{sheng2021phenorm,
  author = {Yu Sheng and Víctor M. Castro and Clara-Lea Bonzel and Molei Liu and Chuan Hong and Tianxi Cai},
  title = {{PheNorm: Unsupervised Gold-Standard Label Free Phenotyping Algorithm for EHR Data}},
  booktitle = {Openalex Unknown},
  year = {2021},
  doi = {10.32614/cran.package.phenorm},
}
Figure