Abstract
As digital platforms like YouTube become primary sources for health information, the gap between medical accuracy and patient understandability grows. This paper proposes an "augmented intelligence" framework designed to curate high-quality healthcare videos, specifically focusing on Type 2 Diabetes. By combining expert medical labeling with semi-supervised co-training, the researchers identify videos that are both medically sound and accessible to patients with varying health literacy levels. Furthermore, the framework audits the corpus for demographic biases in age and gender, revealing significant disparities in representation.
Introduction
While approximately one-third of US adults utilize the internet to diagnose health concerns, only 12% possess proficient health literacy. YouTube's recommendation algorithms frequently prioritize engagement metrics over medical accuracy or "understandability," creating a health literacy divide that disproportionately impacts disadvantaged populations. Addressing this requires a scalable solution that can vet millions of videos while ensuring recommendations are representative and unbiased.
Methodology
The proposed framework utilizes a human-in-the-loop, augmented intelligence approach to synthesize multiple machine learning techniques for video curation and bias auditing.
The process involves:
- Data Collection: A corpus of 9,873 videos related to Type 2 Diabetes was gathered via the YouTube Data API using 200 health-related keywords.
- Semi-Supervised Co-training: To mitigate the need for massive labeled datasets, a co-training model was employed. Two classifiers—one focusing on video metadata (logistic regression) and another on video content (random forest)—iteratively labeled the unlabeled pool based on high-confidence predictions.
- Natural Language Processing (NLP): A Bidirectional Long Short-Term Memory (BiLSTM) model extracted medical terms from descriptions, mapped against the Unified Medical Language System (UMLS).
- Computer Vision for Bias Auditing: FaceLib and pre-trained Caffe models were used to detect faces and estimate the age and gender of presenters to evaluate demographic representation.
Results and Performance
The co-training approach significantly outperformed individual baseline models. The framework achieved an overall accuracy of 87.5% in classifying medical information quality.
Human validation confirmed the model's reliability, with a Kappa score of 0.9 for medical term annotation and 0.87 for video understandability. Statistical analysis using Generalized Linear Models (GLM) indicated that "understandability" was the strongest predictor of video views (GLM = 0.731), suggesting a public preference for accessible content.
Demographic Insights and Bias
The audit revealed a pronounced gender gap within the diabetes video corpus. Most content categorized as having "high understandability" and "high medical information" featured male presenters. While face visibility itself did not significantly impact view counts, the gender of the presenter did (GLM = 0.635), with male presenters garnering more views on average. This suggests systemic biases in how health content is recommended or consumed.
Conclusion and Future Work
This research provides a blueprint for auditing "black box" recommendation algorithms for systemic bias. However, limitations remain, including potential algorithmic bias in the face detection tools themselves and a narrow focus on a single chronic condition. Future work will expand this framework to a broader range of health topics and conduct direct comparisons between "responsible" recommendations and standard platform baselines to further promote health equity.
Summary updated