Concept Graph & Resume using Claude 3 Opus | Chat GPT4 | Gemini Adv | Llama 3:

Resume:

1.-Speaker was surprised by many interesting NLP papers at ICLEAR conference, but most focused on limited topics with ample training data.

2.-Speaker wondered what to do when lacking training data and still wanting good performance, a common issue in medical domains.

3.-Speaker's experience with breast cancer and its treatment revealed machine learning wasn't being utilized much despite ample data.

4.-Only 3% of oncology decisions in the U.S. are based on patients in clinical trials, a small and biased subset.

5.-After treatment, speaker felt compelled to work on impactful problems where NLP could make a difference, despite limited training data.

6.-Collaborating with doctors revealed challenges in extracting information from medical records when lacking labeled data for every condition.

7.-Many medical information extraction systems remain rule-based today due to lack of labeled training data.

8.-Goal became achieving high accuracy with limited supervision doctors could provide in minutes, by transferring knowledge between related tasks.

9.-Approach: Generate task-specific encodings of medical records using limited keyword supervision indicating sentence relevance to the condition.

10.-Adversarial training aligns encodings between source and target tasks to enable a single classifier, along with reconstruction to preserve context.

11.-Experiments showed approach rivaled using in-domain data and outperformed baselines, discovering transferable representations of sentiment across domains.

12.-System is now implemented and used at Massachusetts General Hospital (MGH) to extract breast cancer attributes from pathology reports.

13.-Adding interpretability is key for doctors to trust and utilize machine learning predictions in medical settings.

14.-Goal: Learn to provide extractive rationales for classifications, without rationale annotations, by jointly training generator and predictor.

15.-Beer review experiments showed extractive rationales maintained accuracy while aligning well with human rationales and outperforming baselines.

16.-Rationale system is used by doctors at MGH to quickly see classification explanations, make corrections, and retrain the model.

17.-Speaker realized biggest impact requires going beyond NLP to analyze raw measurements like mammograms to potentially detect early cancer signs.

18.-Studies show potential to identify tissue change patterns in mammograms before cancer appears, possibly preventing cancer with chemoprevention.

19.-First step: Match human radiologist ability to predict breast density and cancer risk (BI-RADS score) from mammograms.

20.-Breast density prediction from mammograms achieved 92% accuracy versus 86% human agreement, a solved task.

21.-BI-RADS score prediction to identify 1% of women needing re-examination was very challenging, with accuracy far below radiologists.

22.-Poor results were corroborated by a recent NYU study, suggesting a very difficult task despite large datasets.

23.-Annotating abnormality locations in a thousand mammograms improved performance to 0.85 AUC, better but still not great.

24.-Speaker hoped to use radiology reports to guide the vision model to abnormalities, but vision model needs improvement first.

25.-Speaker requests ideas from the audience on improving mammogram interpretation to predict cancer risk.

26.-Speaker was motivated by personal experience to work on impactful problems without much labeled data.

27.-Speaker says audience doesn't need to go through a similar ordeal to be motivated to tackle low-resource challenges.

Knowledge Vault built byDavid Vivancos 2024