Every oncologist knows the electronic health record is where the cancer story goes to hide. Outside records arrive by fax, get scanned into the media tab under the wrong name, and pile up to hundreds of documents per patient from dozens of institutions, written in coding languages that never quite translate. Screening one chart for a clinical trial takes an hour. Tumor registries run months behind. Drugs get denied because a performance status was in the note but not in the request. No wonder "data" can feel like a four-letter word. This episode is about what changes when AI can finally read all of it.
In this episode of AI and Healthcare, oncologists Sanjay Juneja, MD and Doug Flora, MD sit down with Mika Newton, CEO of xCures, to unpack a new AI in Precision Oncology paper on using large language models to turn hundreds of medical records per cancer patient into structured, verifiable healthcare data.
In this episode:
• Why a cancer patient's medical records can run to 231 documents from dozens of institutions, and why most of it is still not in discrete fields
• The many "languages" of healthcare data (OMOP, SNOMED, ICD, CPT) and why the crosswalks are never one-to-one
• Learning the structure of a document before extracting from it, and why 95% of the work is organizing rows and columns
• The three buckets: what to automate, where a human belongs in the loop, and when to give up on an unreadable page
• How to vet a data vendor: dual human extraction, model drift, a 17-page buyer's guide, and whether it pays for itself
• Interoperability in healthcare: TEFCA, Individual Access Services, CMS's Kill the Clipboard program, and what changes when patients own their own health data
• The disparity question nobody has solved yet
Mika Newton is Chief Executive Officer of xCures, which has applied AI to organizing and extracting structured data from medical records since 2018. He is a founder of this podcast and a member of the editorial board of AI in Precision Oncology, the journal publishing the paper discussed in this episode. Doug Flora, MD is the journal's Editor-in-Chief; the paper went through independent peer review.
Read the paper: https://journals.sagepub.com/doi/10.1177/2993091X261492372
xCures Buyers Guide: https://xcures.com/interactive-buyers-guide/
Kill the Clipboard: https://www.cms.gov/health-tech-ecosystem/early-adopters/kill-the-clipboard
Subscribe for conversations about artificial intelligence, oncology, clinical medicine, and the future of healthcare.
Hosts:
Sanjay Juneja, MD: https://www.linkedin.com/in/sanjayjunejamd
Douglas Flora, MD: https://www.linkedin.com/in/douglasflora
Follow TensorBlack on X: https://x.com/tensorblack
Follow TensorBlack on Instagram: https://www.instagram.com/tensorblack
Follow TensorBlack on LinkedIn: https://www.linkedin.com/company/tensorblack/
Follow TensorBlack on Facebook: https://www.facebook.com/tensorblack
00:00 AI and Healthcare
00:30 Introduction
01:59 COI
03:01 AI and clinical documentation
04:40 Medicine runs on documents
06:04 OMOP, SNOMED, ICD, CPT: healthcare data languages
08:52 Fax machines and HIPAA
10:04 The media tab problem
11:12 How xCures started
12:03 Document structure before extraction
14:46 231 documents per cancer patient
16:19 Data science is 95% data prep
17:07 Clinical trial screening: an hour per chart
19:14 LLM extraction: enrichment and inference
20:48 AI and healthcare labor
21:58 Human in the loop: the three buckets
24:13 Why the journal published the paper
26:09 AI hallucinations: measure, don't worry
27:17 How to vet a healthcare data vendor
28:05 Dual extraction, model drift, buyer's guide
30:39 Healthcare AI ROI
31:50 EHR integrations vs. patient apps
32:31 Healthcare interoperability, solved in 2023
33:34 TEFCA, Individual Access Services, Kill the Clipboard
35:17 Patient-owned health data
36:00 Health records outside HIPAA
38:22 Digital pathology and health disparities
40:56 Every medical record is a patient
42:47 Data that changes a patient's life
#AIinHealthcare #MedicalRecords #CancerCare
