Predicting Vasovagal Reactions to Needles from Facial Action Units

Judita Rudokaite, Itır Önal Ertuğrul, Lee‐Ling Sharon Ong, Mart P. Janssen, Elisabeth M.j. Huis in ’t VeldJournal of Clinical Medicine 2023 · DOI

TL;DR

This paper introduces a novel machine learning approach to predict vasovagal reactions (VVRs) in blood donors before they donate, by analyzing subtle facial microexpressions captured on video. By leveraging facial action units, the system achieves a high F1 score of 0.82, outperforming traditional self-report methods and offering a promising avenue for proactive intervention.

What Problem Does This Paper Address?

Blood donation, while vital, can trigger unpleasant vasovagal reactions (VVRs) in some donors, ranging from dizziness and sweating to fainting. These reactions are often linked to needle fear and anxiety, which are difficult to self-report accurately and are hard to prevent with current methods. Existing interventions like applied muscle tension or water loading are applied during the donation and may not address the anticipatory anxiety that builds up before the procedure, particularly in the waiting room. This gap leaves a significant portion of donors vulnerable to uncomfortable experiences, potentially impacting future donations and the overall blood supply. The authors aim to bridge this gap by identifying individuals at risk before the needle even appears.

Methods

The core of this study lies in applying machine learning to analyze facial microexpressions. The researchers recorded videos of blood donors in the waiting room and identified 17 specific facial action units (AUs) using the Facial Action Coding System (FACS). They then extracted two types of features from these AUs: their presence (whether an AU was active) and their intensity (how strongly it was expressed).

These AU features were fed into four different machine learning models: decision trees, random forests, XGBoost, and neural networks. To handle the imbalanced dataset (more donors with low VVRs than high), they employed a synthetic minority oversampling technique (SMOTE). Crucially, they used recursive feature elimination with cross-validation (RFECV) for feature selection and GridSearchCV for hyperparameter tuning to build robust models and prevent overfitting. The performance was evaluated using precision, recall, F1 score, and AUC-PR.

Datasets

The study utilized video data from 227 blood donors recruited from three blood collection centers in the Netherlands. These donors were categorized into three groups:

  • Control group: Experienced donors with 5-10 donations and no prior VVRs (n=81).
  • Sensitive group: Donors with 5-10 donations who experienced a VVR at their last donation (n=51).
  • New donors: First-time donors, considered at increased risk (n=95).

The VVR levels were measured using a Likert scale (1-5) for both physiological and emotional symptoms at various stages of the donation process. For model training and evaluation, the sample was split into low VVR (score ≤ 41) and high VVR (score > 41) groups.

Key Insights and Ideas

  • Facial Microexpressions as Predictors: The study's central idea is that unconscious facial movements, even before a needle is seen, can signal an individual's predisposition to VVRs. This moves beyond subjective self-reports to objective physiological cues.
  • Anticipatory Anxiety is Key: The findings highlight that the emotional and physical reactions prior to the actual donation are significant predictors, underscoring the importance of addressing anticipatory fear.
  • Eye and Brow Movements are Crucial: The most predictive features were found in the upper face, specifically related to eye and eyebrow movements. This suggests that subtle changes in how someone looks at or around the needle insertion area can be highly informative.
  • Machine Learning for Proactive Care: The successful application of ML demonstrates its potential to automate the identification of at-risk individuals, enabling targeted interventions before VVRs occur.

Key Results

  • High Predictive Performance: The best-performing model, a neural network utilizing facial action unit intensities with RFECV, achieved an impressive F1 score of 0.82. This significantly outperforms models relying solely on self-reported pre-donation VVR scores (F1 score around 0.77).
  • Recall for High-Risk Donors: The neural network model also demonstrated a strong recall of 0.84 for identifying donors who experienced high VVR levels, meaning it successfully flagged most of those at risk.
  • Most Predictive Features: The intensity of facial action units in the eye regions (e.g., raised upper eyelid, tightened eyelids) and eyebrow movements (e.g., lowered brow) were identified as the most discriminative features for predicting high VVR risk.
  • Group Differences: As expected, new donors and those with a history of VVRs reported significantly higher levels of physiological and emotional reactions compared to the control group.

Limitations and Future Work

  • Cut-off Point: The study acknowledges a limitation in dividing donors into "low" and "high" VVR groups based on a mean cut-off, noting that misclassifications tended to occur around this threshold. Future work aims to model outcomes without a fixed cut-off.
  • Sample Diversity: The authors are collecting more data, particularly from new and "sensitive" donors, to improve model performance and generalizability.
  • Broader Applications: The research team is developing a prototype "AINAR.io" game that uses this biofeedback technology to not only predict VVRs but also to teach individuals how to manage their anticipatory symptoms through a playful interface. This game aims to provide real-time biofeedback, helping users learn self-regulation techniques.
  • Multimodal Sensing: Future work also includes exploring other physiological signals like thermal patterns and heart rate, as well as developing models that directly process video streams without explicit feature extraction.

Summary updated

Abstract

Background: Merely the sight of needles can cause extreme emotional and physical (vasovagal) reactions (VVRs). However, needle fear and VVRs are not easy to measure nor prevent as they are automatic and difficult to self-report. This study aims to investigate whether a blood donors’ unconscious facial microexpressions in the waiting room, prior to actual blood donation, can be used to predict who will experience a VVR later, during the donation. Methods: The presence and intensity of 17 facial action units were extracted from video recordings of 227 blood donors and were used to classify low and high VVR levels using machine-learning algorithms. We included three groups of blood donors as follows: (1) a control group, who had never experienced a VVR in the past (n = 81); (2) a ‘sensitive’ group, who experienced a VVR at their last donation (n = 51); and (3) new donors, who are at increased risk of experiencing a VVR (n = 95). Results: The model performed very well, with an F1 (=the weighted average of precision and recall) score of 0.82. The most predictive feature was the intensity of facial action units in the eye regions. Conclusions: To our knowledge, this study is the first to demonstrate that it is possible to predict who will experience a vasovagal response during blood donation through facial microexpression analyses prior to donation.
BibTeX
@article{rudokaite2023predicting,
  author = {Judita Rudokaite and Itır Önal Ertuğrul and Lee‐Ling Sharon Ong and Mart P. Janssen and Elisabeth M.J. Huis in ’t Veld},
  title = {{Predicting Vasovagal Reactions to Needles from Facial Action Units}},
  journal = {Journal of Clinical Medicine},
  year = {2023},
  doi = {10.3390/jcm12041644},
}
Figure