- The job is reading patient-written drug reviews at scale, not measuring clinical efficacy. The output is a summary of what reviewers say.
- H2O Driverless AI can build text classifiers that label each review effective, ineffective or adverse, with reason codes.
- Reviews that mention a possible adverse reaction go to your safety or medical team for human review. The model never makes a safety determination.
- H2O has no named life sciences customer today. Its proof in this vertical is public competition results, not customer outcomes.
Patients write about your drugs faster than anyone can read
Patients describe their experience with a drug on review sites, forums and blogs. That text tells you what people say about effect, tolerability and side effects, and almost none of it is structured.
Reading it by hand does not scale, and keyword searches miss how people actually write. A post can praise a drug and describe a side effect in the same sentence.
How drug review classification works
H2O can be used to turn unstructured reviews into labeled, searchable records. Your analysts and your safety team decide what the labels mean for your product.
Step 1: collect and clean the review text
Start with the review sources you are permitted to use, plus any labeled examples your team already has. Duplicates, spam and off-topic posts are filtered out before training.
Step 2: turn text into features
H2O Driverless AI applies natural language processing transformers to each review, and feature engineering adds signals such as review length and rating where the source provides them.
Step 3: classify each review
A multiclass classification model labels each review effective, ineffective or adverse. Sentiment analysis adds tone, so a positive review that mentions a side effect is still flagged.
Step 4: route adverse mentions to people
Every review labeled adverse goes to a queue for your safety or medical reviewers, with the passage that triggered it highlighted. Your team decides whether it is a reportable event under your own procedures.
Step 5: document and explain
H2O Driverless AI writes model documentation while it trains, and explainable AI reason codes show which words and features drove each label.
What this is not
H2O makes no claim of GxP, 21 CFR Part 11 or FDA validation. The platform supports validated workflows; the validation is yours. It has no named life sciences customer today.
Proof
H2O can be used for review classification and adverse-mention routing, though it has no named life sciences customer today. The modeling proof in this vertical is public and checkable: first place in the Kaggle diabetic retinopathy detection competition, first of 784 entries in a Kaggle pulmonary embolism computer-vision challenge, and top three in the HEROHE histopathology challenge. These are competition results, not customer outcomes.