Mining clinical text notes to facilitate surveillance of Alabama Rot

Jenny Newman, Pj Noble, Goran Nenadić, Alan Radford, David Walker, Jessica Bacon, Philip H. JonesBritish 2015 · DOI · OpenAlex

Abstract

Canine idiopathic cutaneous and renal glomerular vasculopathy (CRGV), popularly known as "Alabama Rot," is a rare and devastating disease that emerged in the UK in 2012. Due to its rarity and the lack of a definitive ante-mortem diagnostic test, it is frequently under-diagnosed. This paper details a research effort to develop a text-mining and machine learning pipeline capable of identifying CRGV cases within unstructured veterinary clinical notes. By analyzing "sentinel features"—linguistic patterns in symptom descriptions—the researchers built a model that achieved 100% sensitivity on confirmed cases and identified dozens of potential undiagnosed cases in a large control group. This approach offers a scalable method for real-time surveillance of emerging veterinary diseases.

Introduction

The emergence of Canine idiopathic cutaneous and renal glomerular vasculopathy (CRGV) in the UK presented a significant challenge to veterinary medicine. Characterized by skin lesions and acute kidney injury, the disease is often fatal. Traditional surveillance relies on clinicians manually reporting suspected cases, a process that is inherently slow and prone to under-reporting.

To address this, researchers investigated whether the unstructured text within electronic health records (EHRs) could serve as a "linguistic biomarker." By identifying specific combinations of clinical descriptions that correlate with confirmed CRGV, it is possible to automate the detection of potential cases, providing a safety net for clinicians and a more accurate picture of disease prevalence.

Methodology

The study utilized a multi-stage Natural Language Processing (NLP) and machine learning workflow to process clinical data from the Small Animal Veterinary Surveillance Network (SAVSNET).

Data Processing and Feature Selection

The pipeline began with text cleaning and semantic clustering, where words with similar meanings were grouped together. To identify the most predictive features, the researchers calculated the likelihood of specific word clusters appearing in confirmed cases versus a control group. To ensure a high signal-to-noise ratio, only clusters with an odds ratio > 20 and a p-value < 0.0001 were retained.

Model Development

Clinical records were converted into binary vectors representing the presence or absence of these high-significance "key entity groups." The researchers then used the WEKA (Waikato Environment for Knowledge Analysis) workbench to build and optimize a predictive model. The goal was to find the optimal balance between sensitivity (catching all true cases) and specificity (minimizing false alarms).

Datasets

The model was trained and tested using two primary groups:

  • Case Group: 33 dogs with histologically confirmed CRGV.
  • Control Group: 3,300 dogs randomly selected from the SAVSNET database, representing general consultations for illness or injury between September 2012 and January 2013.

Results and Discussion

The final model demonstrated exceptional performance, achieving 100% sensitivity by successfully identifying all 33 confirmed cases of CRGV.

Beyond the known cases, the model flagged 58 out of the 3,300 control animals as matching the linguistic profile of the disease. While these require further clinical investigation to rule out false positives (such as other forms of acute kidney injury), they likely represent a pool of previously undiagnosed cases. This suggests that the true prevalence of Alabama Rot may be higher than manual reporting indicates.

Clinical Implications

This research proves that even with a small sample of positive cases, statistically rigorous text mining can extract highly predictive features from noisy clinical data. The resulting tool acts as an automated differential diagnosis aid, prompting clinicians to consider rare diseases when a patient's clinical notes match a known "fingerprint."

Conclusion and Future Directions

The success of this NLP pipeline in identifying Alabama Rot cases highlights the potential for large-scale, automated surveillance of emerging veterinary syndromes. Future work will focus on validating the model against independent datasets and refining the analysis of flagged control cases to improve specificity. This methodology is broadly applicable and could be adapted to monitor a wide range of rare or emerging diseases across the veterinary sector.

Summary updated

Abstract

Preview this chapter: Mining clinical text notes to facilitate surveillance of Alabama Rot, Page 1 of 1 < Previous page | Next page > /docserver/preview/fulltext/10.22233/9781910443521/9781910443521.63.6-1.gif
BibTeX
@inproceedings{newman2015mining,
  author = {Jenny Newman and PJ Noble and Goran Nenadić and Alan Radford and David Walker and Jessica Bacon and Philip H. Jones},
  title = {{Mining clinical text notes to facilitate surveillance of Alabama Rot}},
  booktitle = {British},
  year = {2015},
  doi = {10.22233/9781910443521.63.6},
}
Figure