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

Area of Science:

• Clinical chemistry and laboratory medicine
• Artificial intelligence in healthcare
• Statistical modeling for biological data

Background:

• Reference interval (RI) estimation is often inaccurate with small sample sizes.
• Visual assessment of distribution histograms (VADH) can aid statistical technique selection but relies on human interpretation.
• Developing automated methods for VADH is crucial for improving RI accuracy in limited data scenarios.

Purpose of the Study:

• To develop and validate a machine learning model for automated visual assessment of distribution histograms (VADH) in small sample sizes.
• To compare the performance of the machine learning model against traditional statistical tests for predicting population distributions.
• To create a user-friendly web application for applying the developed model in clinical practice.

Main Methods:

• A convolutional neural network (CNN) was trained on 45,000 simulated distribution histograms from various population types with sample sizes ranging from 20 to 40.
• The CNN model's accuracy in predicting original population distributions was evaluated on a test set of 900 human-classified histograms.
• Performance was benchmarked against the Shapiro-Wilk test, and a web application (HARISS) was developed for practical implementation.

Main Results:

• The CNN model achieved high accuracy in predicting population distributions by VADH, reaching 84.0% on the training set and 94.4% on the test set.
• In comparison, the Shapiro-Wilk test showed lower accuracy, with 65.0% and 72.3% on the test set using different p-value thresholds.
• The HARISS web application was successfully deployed, offering VADH, RI estimation, and outlier detection functionalities.

Conclusions:

• The developed CNN model effectively performs VADH, offering a significant improvement over manual assessment and traditional statistical methods for small sample sizes.
• The HARISS web application provides a practical tool to enhance the accuracy of reference interval estimation.
• While the model shows promise, careful selection of reference individuals and adherence to preanalytical/analytical conditions remain critical for accurate RI estimation.