Abstract
As YouTube becomes a primary source for health information, the challenge of distinguishing high-quality educational content from low-value "noise" grows. This paper presents a machine learning pipeline designed to automatically identify and curate medically accurate, understandable health information. By combining natural language processing (NLP) and computer vision features, the system achieves over 83% accuracy in recommending videos suitable for public health literacy, using COVID-19 information as a primary case study for validation.
Introduction
While YouTube is the world's largest repository of health information, it suffers from a significant signal-to-noise problem. For the general public—particularly those with lower health literacy—purely text-based medical instructions are often difficult to parse. Video offers a more engaging medium; however, platform algorithms often prioritize popularity metrics (views and engagement) over medical accuracy or educational clarity. This research addresses the challenge of "curation at scale," proposing an automated method to shortlist videos that are both medically sound and follow best practices for patient education.
Methodology
The researchers developed a multi-stage pipeline to bridge the gap between raw video data and expert-level curation:
- Keyword Generation: A triangulation approach utilized patient forums (DailyStrength), YouTube search suggestions, and official FAQs from the WHO and CDC to identify common user queries.
- Feature Extraction: Using the YouTube Data API and Google Cloud Video Intelligence API, the system extracted metadata, speech-to-text transcriptions, and Optical Character Recognition (OCR) data from on-screen text. Visual features, such as the number of "shots" (scene changes), were used as proxies for production quality.
- Medical Entity Recognition: A Bidirectional Long Short-Term Memory Recurrent Neural Network (BLSTM-RNN) paired with the Unified Medical Language System (UMLS) was employed to quantify the density of medical information.
- Understandability Scoring: The Patient Education Materials Assessment Tool for Audiovisual Materials (PEMAT-A/V) framework provided a standardized metric for how easily a video could be followed by a layperson.
Key Results
The study evaluated four machine learning models—Logistic Regression, Support Vector Machines (SVM), Random Forest, and XGBoost—to predict whether a video should be recommended based on human-labeled ground truth. Logistic Regression emerged as the most effective model for this task.
As shown in the results, Logistic Regression achieved an accuracy of 83.61% and a recall of 75.86%, outperforming more complex ensemble methods like XGBoost on the curated dataset of 304 videos.
Discussion and Insights
A critical finding of this research is that popularity does not equate to quality. YouTube’s internal ranking system often highlights viral content that may lack medical relevance or clarity. The most predictive features for a "recommended" video were the understandability score, medical information density, and the readability of the video description.
Recommended videos typically feature factual, timely guidance provided by medical professionals and are supported by visual aids like slides. In contrast, non-recommended videos often focus on news or political updates with minimal medical information, despite high view counts.
Conclusion
This automated pipeline offers a scalable solution for public health organizations to curate high-quality video content. By prioritizing understandability and medical density over simple engagement metrics, the system can significantly improve the quality of health information available to the public. Future work will focus on incorporating "timeliness" as a dynamic feature and validating these curated playlists through randomized controlled trials to measure actual patient health outcomes.
Summary updated