Leveraging Generative AI and Transformers in Healthcare
summary
An anomaly-identification pipeline ingests social media and email data, labels ~20,000–30,000 records, fine-tunes BERT/BioBERT/XLM‑RoBERTa via Hugging Face, tracks experiments with MLflow and serves models with PyTorch; an ensemble plus heuristic guardrails flags suspicious posts and slashes manual review from millions to a manageable queue, targeting 98–99% recall to avoid missed events, and evaluates using GPT‑4/Llama 2 with prompt engineering to perform classification without fine‑tuning.
AI-written summary · about
chapters (14)
- 0:00intro
- 0:20preamble
- 1:05anomaly detection
- 2:27the challenges we set out to solve
- 4:51anomaly detection at a glance
- 5:47ml algorithm development process
- 7:20why data annotation?
- 8:24data annotation...
- 10:04a comprehensive ai impl journey
- 12:19anomaly identification framework
- 14:29machine learning model framework - continuous training for optimized performance
- 16:33llm + guardrails
- 17:18challenges in bert models
- 19:35generative ai and prompt engineering
related talks
18mMedical AI Beyond Prediction: From Normative Learning to Agentic Healthcare AI
22mLeveraging AI/ML for Transformative Software Solutions in Healthcare
16mGenAI in Healthcare Claims: Processing and Detecting Fraud
33mLessons Learned Applying Large Language Models in Healthcare
41mMerging Human Cyber Security Expertise with Generative AI
26mNatural Language Processing for Healthcare: Current Trends...
24mEngaging new and diverse audiences with ML
35mGenerative AI for Fraud Detection & Claims Analysis