- A classification model scores each new sample from its measurements or images and shows the factors behind the score, so your lab can decide which slides to review first.
- Enterprise h2oGPTe reads free-text pathology reports and returns a structured summary of findings, with the passage behind each finding cited.
- This is a review-ordering and evidence-finding tool, not a diagnosis: a pathologist reviews and signs every result.
- Protected health information never has to leave your boundary across on-premises, air-gapped, private cloud and cloud deployment.
Every sample gets the same place in the queue
Your cytopathology lab screens a steady flow of samples, including liquid biopsies, and every one gets the same place in the queue. The samples that need a second look wait behind the ones that do not.
The knowledge your team needs is already written down. It sits in thousands of free-text pathology reports, and finding a pattern across them means reading them one by one. Anything you add also has to fit the lab's workflow and quality process, keep patient data inside your environment, and leave the diagnosis with the pathologist.
How cytopathology AI works
Cytopathology AI combines two steps: a classification model scores each new sample from its measurements or images, and a retrieval layer turns your report archive into answers your team can check.
Step 1: start from reviewed cases
Use samples your pathologists have already signed out, with their measurements, images and final categories.
Step 2: build and compare models
H2O Driverless AI builds and ranks scoring models on your data. Unsupervised machine learning and similarity search group samples that look alike.
Step 3: explain each score
ML interpretability shows which measurements drove each score, so a reviewer can check it against the slide.
Step 4: sort the queue
Your lab's policy decides how scores reorder review. Nothing is signed out without a pathologist.
Step 5: summarize reports with citations
Enterprise h2oGPTe uses retrieval-augmented generation to extract findings from your pathology reports and cite the passage behind each one.
What the model does not decide
H2O makes no clinical efficacy claim for this use case. The model changes where your pathologist starts and gives them cited findings from past reports; it does not make the diagnosis, and every sample is still signed out by a person.
Most labs start in shadow mode, running the model beside the current review order and comparing before anyone acts on it, so the queue change is tested before it is trusted.