Discriminant analysis to predict the clinical diagnosis of primary immunodeficiencies: a preliminary report

Abstract

Background: The features in a clinical history from a patient with suspected primary immunodeficiency (PID) direct the differential diagnosis through pattern recognition. PIDs are a heterogeneous group of more than 250 congenital diseases with increased susceptibility to infection, inflammation, autoimmunity, allergy and malignancy. Linear discriminant analysis (LDA) is a multivariate supervised classification method to sort objects of study into groups by finding linear combinations of a number of variables.

Objective: To identify the features that best explain membership of pediatric PID patients to a group of defect or disease.

Material and method: An analytic cross-sectional study was done with a pre-existing database with clinical and laboratory records from 168 patients with PID, followed at the National Institute of Pediatrics during 1991-2012, it was used to build linear discriminant models that would explain membership of each patient to the different group defects and to the most prevalent PIDs in our registry. After a preliminary run only 30 features were included (4 demographic, 10 clinical, 10 laboratory, 6 germs), with which the training models were developed through a stepwise regression algorithm. We compared the automatic feature selection with a selection made by a human expert, and then assessed the diagnostic usefulness of the resulting models (sensitivity, specificity, prediction accuracy and kappa coefficient), with 95% confidence intervals. 

Results: The models incorporated 6 to 14 features to explain membership of PID patients to the five most abundant defect groups (combined, antibody, well-defined, dysregulation and phagocytosis), and to the four most prevalent PID diseases (X-linked agammaglobulinemia, chronic granulomatous disease, common variable immunodeficiency and ataxia-telangiectasia). In practically all cases of feature selection the machine outperformed the human expert. Diagnosis prediction using the equations created had a global accuracy of 83 to 94%, with sensitivity of 60 to 100%, specificity of 83 to 95% and kappa coefficient of 0.37 to 0.76. 

Conclusions: In general, the selection of features has clinical plausibility, and the practical advantage of utilizing only clinical attributes, infecting germs and routine lab results (blood cell counts and serum immunoglobulins). The performance of the model as a diagnostic tool was acceptable. The study’s main limitations are a limited sample size and a lack of cross validation. This is only the first step in the construction of a machine learning system, with a wider approach that includes a larger database and different methodologies, to assist the clinical diagnosis of primary immunodeficiencies.