YouTube Video Analytics for Patient Engagement: Evidence from Colonoscopy Preparation Videos

Yawen Guo, Xiao Liu, Anjana Susarla, Rema PadmanarXiv 2024 · DOI · arXiv

Introduction

Patient education is a cornerstone of effective healthcare, particularly for procedures like colonoscopy that require complex home preparation. While YouTube offers a vast repository of educational content, it remains a "Wild West" of information, ranging from high-quality professional guidance to dense academic lectures and potentially misleading user-generated content. Manually vetting these millions of videos for both medical accuracy and patient understandability is an impossible task for clinicians at scale.

This research introduces an automated pipeline designed to evaluate the quality and accessibility of medical education videos. By leveraging deep learning and multimodal feature extraction, the authors developed a system capable of identifying patient-appropriate content with high precision, using colonoscopy preparation as a primary case study.

Methodology and Data

The researchers developed a multi-stage analysis pipeline to process 312 English-language YouTube videos related to colonoscopy. Keywords for video discovery were sourced from the DailyStrength health community, search suggestions, and existing academic literature.

The pipeline consists of four main components:

  1. Feature Extraction: Utilizing the YouTube Data API and Google Video Intelligence API to gather metadata and multimodal signals, including Optical Character Recognition (OCR) for on-screen text, label detection for visual activities, and speech-to-text transcripts.
  2. Medical Entity Recognition: To identify relevant medical terminology, the team implemented a Bidirectional Long Short-Term Memory (BiLSTM) neural network. This model was trained on 19,300 sentences to recognize terms from the Unified Medical Language System (UMLS).
  3. Expert Annotation: A subset of videos was manually annotated by researchers using the Patient Education Material Assessment Tool (PEMAT) framework to establish ground truth for medical information levels and understandability.
  4. Classification: Logistic regression models were trained to categorize videos based on their suitability for patient recommendation.

Results and Performance

A key technical contribution of this work is the application of BiLSTM for medical entity extraction. The BiLSTM model significantly outperformed the traditional Conditional Random Field (CRF) baseline, capturing the nuanced context of medical terms more effectively. This contextual awareness is vital in health communication, where the surrounding words can fundamentally change the meaning of a procedure or symptom.

Performance comparison of CRF and BLSTM RNN models for medical entity extraction.

The final "Video Recommendation" classifier achieved an impressive overall accuracy of 94.7%. Analysis of the predictive features revealed that multimodal signals are critical indicators of quality. Specifically, "OCR confidence" (the clarity of text on screen) and the "number of shots" (visual pacing) were strong predictors of a video's educational value. Recommended videos tended to have higher transcription confidence and a higher density of unique medical terms, provided they remained readable.

Descriptive statistics for video-level and metadata features used in the classification models.

Discussion and Future Directions

The study highlights a common "Expert Trap" in online health content: many videos contain high-density medical information but suffer from low understandability (e.g., recorded medical school lectures). The developed pipeline successfully distinguishes these from patient-centered content by analyzing readability scores and visual production quality. This suggests that well-produced videos with clear visual aids are fundamentally more "recommendable" to patients than static recordings.

While promising, the study is limited by its sample size and focus on a single procedure. Future work aims to expand this framework to other chronic conditions and incorporate direct feedback from clinicians and patients to refine understandability metrics. Additionally, the authors plan to investigate the causal impact of these recommended videos on actual patient compliance and health outcomes, such as improved adherence to colonoscopy preparation instructions.

Summary updated

BibTeX
@misc{guo2024youtube,
  author = {Yawen Guo and Xiao Liu and Anjana Susarla and Rema Padman},
  title = {{YouTube Video Analytics for Patient Engagement: Evidence from Colonoscopy Preparation Videos}},
  year = {2024},
  eprint = {2410.02830},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.02830},
}
Figure