Detection of Alcohol, Drug, and Sleepiness Conditions through Iris Behavior Curve Analysis

Abstract

This paper provides an extensive analysis of iris behavior analysis as a means of determining fitness-for-duty by identifying impairment stemming out of alcohol, drugs and sleep deprivation. The system utilizes near-infrared image sequences to obtain biometric characteristics whereby the pupil and iris radii, their ratios and spatial positions are obtained and indicate substantial statistical variations between the control and impaired groups. The power of these features is confirmed by Kruskal-Wallis and Dunn tests, where the H-statistics are over 1200 of the important variables and p-values are below 0.001 in most the group comparisons, which prove that the selected biometric markers are discriminatory under different conditions. Models of classification such as the Random Forest, Gradient Boosting and the Multi-Layer Perceptron are reported to be having excellent accuracy up to 75.5 percent with the sensitivity and specificity of the model often being above 70 percent and 95 percent respectively. It is interesting to note that binary classification with fit and unfit states consistently has better performance and often has a sensitivity over 80% to detect unfit persons and therefore the practical use of the system in the occupational safety field. The non-invasive nature of the approach used by the system allows continuous monitoring without interrupting the work of operators: driving or piloting, which is why there are only fewer cases of sleep and drug impairment affecting the sensitivity of the classification in these groups. Segmentation can also be affected by image quality and limitations of NIR sensor. Future directions in work will aim to add to the data set with attention paid to underserved conditions, and the incorporation of deep sequential models like LSTMs to better model the temporal dynamics of biometric. Additional refinement of image processing, sensor strength, and ability to execute it in real-time will increase accuracy as well as application utility Such developments have scalable proactive safety evaluation potential, especially in safety-critical sectors that use biometric indicators to conduct real-time detectors of impairments.