Combating PTSD with Machine Learning
summary
Claims backlog exceeds 300,000 with PTSD making ~22% and appeal timelines over four years, and automation is proposed to speed adjudication. Using a 6,153-sentence dataset of BVA decisions the pipeline classifies six rhetorical roles like FindingSentence and ReasoningSentence. Initial SVM hits 86% accuracy, a CNN+LSTM drops to 76%, and the final model uses GloVe-300 embeddings, 128-unit bidirectional LSTM, 256 dense layer, 0.5 dropout and softmax to reach 91% accuracy with 99% precision on citations.
AI-written summary · about
chapters (10)
- 0:00Start
- 2:03What is Machine
- 3:00Veterans' Claim Appeals
- 3:57Research Goals
- 5:18Key Components
- 6:07JSON Dataset
- 6:29Sample Sentences
- 7:33Five Machine Learning Concepts
- 8:14Deep Learning with Python by François Chollet
- 8:36Definition: SVM creates a line or a hyperplane which separates the data into classes.
related talks
31mWhen humans teach machines
25mMachine Learning Spacecraft Designing for Cybersecurity
20mAugmenting Mental Health Care in the Digital Age: Machine Learning as a Therapist Assistant
01mMachine Learning for Human Rights Advocacy
29mDetecting anomalous sequences using text processing methods
28mMachine Learning applied to judicial decisions
18m: Workforce Efficiency with AI-Driven Vector Search
34mClassifying Search Queries Without User Click Data