Computational diagnosis and risk evaluation for canine lymphoma

Eugenij Moiseevich Mirkes, I. Alexandrakis, K. Slater, R. Tuli, Alexander N. GorbanComputers in Biology and Medicine 2014 · DOI · OpenReview

TL;DR

This paper presents a computational system for diagnosing and assessing the risk of canine lymphoma using a blood test that measures two acute phase proteins (CRP and Haptoglobin). The developed system achieves high sensitivity and specificity for both differential diagnosis and screening, and importantly, can detect lymphoma recurrence up to two months before clinical signs appear.

What Problem Does This Paper Address?

Canine lymphoma is a common cancer in dogs, and early and accurate diagnosis is crucial for effective treatment and improved outcomes. Current diagnostic methods can be invasive or lack the sensitivity needed for early detection and monitoring. The paper addresses the need for a non-invasive, reliable, and computationally driven approach to diagnose canine lymphoma, screen for it in at-risk populations, and monitor for recurrence after treatment. This is particularly challenging due to the heterogeneity of the disease and the need to distinguish it from other conditions.

Methods

The authors employed a range of machine learning techniques to build their diagnostic system. They focused on three main families of methods:

  • Decision Trees: Various algorithms were tested, including those based on information gain, Gini gain, and a custom criterion called DKM gain. They explored different impurity criteria, kernel functions, and feature sets.
  • K-Nearest Neighbors (KNN): This included standard KNN, as well as advanced and adaptive versions. They experimented with different distance metrics (Euclidean, Fisher's transformed, adaptive) and kernel functions.
  • Probability Density Function Estimation (PDFE): Radial basis functions were used for this approach, allowing for direct estimation of lymphoma risk.

Several pre-processing techniques were also investigated, including the use of logarithmic transformations for protein concentrations and feature selection. The evaluation of these methods was primarily done using Leave-One-Out Cross-Validation (LOOCV) to ensure robust performance estimates.

Datasets

The study utilized a database of 303 canine records. These records included:

  • Categorical Input Features: Sex, Lymphadenopathy, Neutered, Breed.
  • Real-Valued Features: Age, concentrations of Haptoglobin (Hapt), and C-Reactive Protein (CRP).

The data was sourced from dogs undergoing differential diagnosis for lymphoma and from healthy dogs obtained from a pet blood bank. The database was analyzed to identify two distinct cohorts: "clinically suspected" cases and "healthy" cases, which informed the formulation of two distinct problems: differential diagnosis and screening.

Key Insights and Ideas

  • Biomarker Synergy: The combination of CRP and Haptoglobin levels, along with clinical symptoms like lymphadenopathy, proved to be a powerful predictor of lymphoma.
  • Early Recurrence Detection: The developed system can detect lymphoma recurrence up to two months before clinical signs become apparent, offering a significant advantage for timely intervention.
  • Risk Mapping for Hypothesis Generation: Visualizing the risk of lymphoma as a "risk map" based on CRP and Haptoglobin levels provides an intuitive tool for understanding feature impact and generating hypotheses about disease drivers.
  • Importance of Pre-processing: Logarithmic transformation of protein concentrations and careful feature selection were crucial for optimizing model performance.
  • Decision Trees as a Strong Performer: Decision trees, particularly those using the DKM splitting criterion, emerged as the best-performing models for both differential diagnosis and screening tasks.

Key Results

  • Differential Diagnosis: The best solution, a decision tree using CRP, Haptoglobin, and Lymphadenopathy, achieved a sensitivity of 83.5% and a specificity of 77% using LOOCV.
  • Screening: For screening, a decision tree using the same three input features achieved a sensitivity of 81.4% and a specificity of >99%. When Lymphadenopathy was considered unknown, a decision tree using only CRP and Hapt provided a sensitivity of 69% and a specificity of 83.5%.
  • Recurrence Monitoring: The system demonstrated the ability to detect lymphoma recurrence up to two months prior to the appearance of clinical signs.
  • Risk Evaluation: The developed models provide probabilistic risk estimates, enabling more informed decision-making.

Limitations and Future Work

  • Breed Inclusion: The authors noted that the large number of breeds and limited data per breed prevented the inclusion of "Breed" as a direct feature, suggesting that breed clustering or other numerosity reduction techniques might be beneficial for future work.
  • Data Imbalance for Screening: For screening, particularly for dogs without obvious lymphadenopathy, the authors suggest further clinical testing is needed, especially for cases with lymphoma but no visible symptoms, as there were only four such cases in the dataset.
  • Future Directions: The authors propose several avenues for future research, including:
  • Further clinical validation of the screening classifier, especially for cases without obvious lymphadenopathy.
  • Continued testing of the treatment monitoring system to confirm its impact on treatment planning and survival rates.
  • Investigating breed clustering for improved diagnostic models.
  • Exploring combinations of CRP and Hapt with results from routine blood tests for optimal feature sets.

Summary updated

BibTeX
@article{mirkes2014computational,
  author = {Eugenij Moiseevich Mirkes and I. Alexandrakis and K. Slater and R. Tuli and Alexander N. Gorban},
  title = {{Computational diagnosis and risk evaluation for canine lymphoma}},
  journal = {Computers in Biology and Medicine},
  year = {2014},
  doi = {10.1016/j.compbiomed.2014.08.006},
}
Figure