Healthcare AI Daily Briefing

Industry & Products

Four AI levers could add ~$125k revenue per physician in specialty practices.

Prosper AI studied call and claims data from more than 2,000 practices across 29 states and six specialties, finding that four AI‑driven levers—improving call answer rates, patient follow‑up, pre‑visit benefit checks, and patient‑responsibility recovery—together yield an estimated $125,000 in extra annual revenue per physician. [5]

Operations & Workflow

Jefferson Health targets ten million hours returned to care in three years.

Podcast from Claims Denied: A Hospitalogy Podcast

The interview describes Jefferson Health’s goal to return ten million hours to care in three years, with its first tranche of AI tools focused on reducing documentation time for nurses, doctors, pharmacists and after‑hours staff. It notes the AI center of excellence, established about three years ago, evaluates and rolls out the tools. [4]

Medical Imaging

Sopris Capital buys Bayer’s radiology AI unit Blackford Analysis.

HIStalk reported that Sopris Capital acquired Blackford Analysis, the AI medical imaging business Bayer announced it was shutting down. The deal transfers Blackford’s radiology AI portfolio to Sopris, which plans to integrate it with its own holdings, though financial terms were not disclosed. [6]

Research & Evidence

Developer‑recommended input alterations raise Jev staging accuracy to 85.8%.

Preprint

In a medRxiv preprint, the decision model Jev achieved 45.4% accuracy for ordinal colon cancer staging with standard LLM‑style inputs, below the 94.2‑100% range of five flagship LLMs; after developer‑recommended input alterations (medical context + task decomposition) its accuracy rose to 85.8%, and none of Jev’s staging errors were made with full confidence. [2]

Clinical Practice

XGBoost predicts pacing‑induced cardiomyopathy with AUROC 0.831 from internal cross‑validation.

Preprint

In a multicenter prospective study of 125 patients with atrioventricular block and HFmrEF/HFpEF, four ML algorithms were trained using stratified 4‑fold cross‑validation; XGBoost showed the highest discrimination for pacing‑induced cardiomyopathy (AUROC 0.831, sensitivity 0.805, specificity 0.845) from this internal validation, with baseline LVEF and post‑implant QRS duration as the most important features. [1]

Industry & Products

Interview: AI could automate about 50% of documentation tasks.

Podcast from AI and Healthcare

The interviewee said AI could automate about 50% of documentation tasks because they are the easier ones on a scale of easiness, and would generate a signal to push those cases through automation. [3]