(P-BC-66) Utilization of Artificial Intelligence (AI) in evaluating emotional aspects of blood donors – a novel concept in predicting Adverse Donor Reaction (ADR) – Developing AI based model

Enhancing the safety and sustainability of blood donation is a paramount concern, with voluntary donors serving as a vital source of "safe blood". Around 10%-12%, of blood donors experience ADRs, with nearly 9% of these individuals opting out of subsequent donations. Retaining repeat voluntary donors necessitates minimizing the occurrence of ADRs. Emerging field of AI, based on the principle of learning, reasoning and self correction results in better predicting the outcomes. AI with its Machine Learning (ML), deep learning capacities, has been used in Transfusion Medicine. We have used AI based model to evaluate real time emotions from the facial expressions of the blood donors and analyzed to determine whether the donors would have ADR. We are developing this model for accurate predictions of ADR.

Study

Design/Methods:

Retrospective analysis: The donor is screened by self check-in process using a TABLET wherein a selfie  is taken. In the first stage of model development, the photos were evaluated and analyzed for emotions such as Neutral, fear, sadness, contempt, happiness, surprise  and disgust to determine  correlation with ADR. They were analyzed on a scale of 0 to1 where 0 was normal and 1 was extreme. Analysis was done using Microsoft cognitive Intelligence tools. Since the correlation was established, it was decided to build a model to proactively predict ADR. The second stage was building the ML model and training of an historical data and checking the accuracy on Test data. This was done using XG Boost along with upscaling (SMOT–Synthetic Minority Over sampling Technique) and hyper tuning. Since the model was predicting normal donors but was not correctly predicting donors with ADR, in the third stage a second image of the donor during  phlebotomy was incorporated and data was analyzed with both the images.

Results/Findings:

In the first stage, 2332 donors were analyzed out of which 294 had ADR. Analysis of the images of the donors who had ADR showed that contempt and sadness were the major indicators. Surprisingly, happiness specifically when reading was high on the scale ( >0.9), was also an indicator. In the second stage, 2806 donors were analyzed, 2663 were normal , 37 had ADR, 4 were correctly predicted by the model (10.8%). In Third stage, 2366 donor images were analyzed (1^st and 2^nd image model). 2334 were normal and model predicted 14/32 to have ADR (43%).

Conclusions:

Emotions of the donors are correlated with the occurrence of ADR. The AI based model predicted normal donors with good success rate, however prediction of ADR needs improvement. Addition of second donor image during phlebotomy has improved the prediction rate of ADR. A pilot study, being AI based model, more and more data will be analyzed for improving the prediction of ADR.