HARISS: Histogram Analyzer for Reference Intervals of Small Samples, a Free Web App to Calculate Reference Intervals of Small Samples.

Abstract

BACKGROUND: Reference interval (RI) estimate inaccuracy is problematic at small sample sizes. Visual assessment of distribution histograms (VADH) may improve statistical technique selection, but its performance depends on the human operator. OBJECTIVES: This study aimed to develop a machine learning model for small-sample VADH. METHODS: The training set consisted of 45 000 distribution histograms from samples ranging from 20 to 40 individuals extracted from simulated Gaussian, lognormal, and left-skewed populations. A convolutional neural network (CNN) was trained to predict the original population distribution by VADH. Its accuracy was tested on 900 human-classified histograms from a previous study (ranging from 20 to 60 individuals) and compared to the Shapiro-Wilk test performance in determining the original population distribution. A web application was developed to allow usage of the CNN, 95% RI estimation with 90% confidence intervals (CI) via bootstrapping, and outlier detection. RESULTS: The CNN model properly predicted the original population distribution by VADH in 84.0% (95% CI: 83.7-84.4) and 94.4% (95% CI: 92.7-95.8) of the samples of the training and the test sets, respectively. Comparatively, on the test set, the Shapiro-Wilk test accuracy was 65.0% (95% CI: 61.8-68.1) and 72.3% (95% CI: 69.3-75.2) using a p-value threshold of 0.05 and 0.2, respectively. The web application (named HARISS) was successfully deployed and is accessible at: https://hariss.streamlit.app/. CONCLUSIONS: The CNN model demonstrated effective VADH and might enhance RI estimate accuracy through the use of the HARISS web app, though proper reference individual selection and preanalytical/analytical conditions remain paramount.