Home › Blog › Oncology › PRECISION-AI Study

PRECISION-AI Study: Key Findings

• Study: Three oncologists at three tertiary oncology centres in India rated 110 clinical queries across five cancer types.
• Accuracy: 82.7% of answers rated fully correct, with a composite rate of 100%.
• Safety: 100% of answers rated clinically safe.
• Guideline concordance: 94.5% of answers rated fully concordant, with a composite rate of 100%.
• Time: The median time to a clinical recommendation fell from 389.2 seconds to 47.2 seconds.

Why Medical AI in Oncology Needs Testing in India

India records an estimated 1.4 million new cancer cases every year, and the paper cites projections that incidence will double by 2040. The oncology workforce has not kept pace. The authors put the ratio at about one oncologist for every 1,600 new diagnoses a year, with the widest gaps in smaller cities.

Any oncologist running a full outpatient clinic knows what that ratio means in practice. Guidelines from NCCN, ESMO, ASCO and ICMR are revised regularly, and the evidence behind them grows every month. Checking a plan against the current version takes several minutes, and a crowded clinic rarely offers them.

AI tools that summarise medical evidence have spread quickly, and many clinicians already try general-purpose chatbots for a first look. Whether a given tool has been tested by the doctors who would rely on it, in the setting where it would be used, is a separate matter. The authors note that earlier evaluations each covered only part of the picture, such as accuracy without safety, or perceived usefulness without guideline concordance. PRECISION-AI was designed to measure these together, in Indian practice.

What Was Evaluated: The OneRx Medical AI Platform

The study evaluated OneRx, written as OneRx.AI in the paper. OneRx is a medical AI platform for evidence summary for clinicians. When a doctor asks a clinical question, the platform retrieves the relevant content from current NCCN, ESMO, ASCO and ICMR guidelines and builds its answer from that material, with citations the doctor can check.

The purpose of this design is to keep each answer tied to a source that can be verified, in place of relying only on what a language model absorbed during training.

The paper is titled “Prospective multicentre expert evaluation of an artificial intelligence-based oncology clinical decision support system across major cancer types in India.” It appears in Volume 222 of the International Journal of Medical Informatics, a peer-reviewed journal published by Elsevier. PRECISION-AI is a short study name and not a formal acronym.

A note on who is writing: this article is published by OneRx, the platform that was evaluated. The paper reports that the study had no funding source and that the authors declared no competing interests. The authors thank the OneRx medical affairs and technical teams for arranging access to the platform for research.

How the PRECISION-AI Study Was Designed

Three oncologists across three independent tertiary oncology centres took part. Data collection ran for 15 days within a 45-day study period in March and April 2026. The work had two phases, which were analysed separately.

1. Individual evaluation, days 1 to 10. Each oncologist put real-world clinical queries to the platform and rated the answers independently.
2. Common prompt evaluation, days 11 to 15. The oncologists rated a shared set of prompts, so that agreement between them could be measured.

Across the three centres, 124 queries were submitted and 110 (88.7%) met the eligibility criteria. All queries were deidentified, and the paper states that no patient-identifiable data was collected at any stage.

Cervical or uterine cancers made up the largest group, with 46 queries (41.8%). Head and neck or oral cancers followed with 39 queries (35.5%). Breast cancer accounted for 12 queries (10.9%), lung cancer for 7 (6.4%) and colorectal cancer for 6 (5.5%).

All five are among the cancers that account for most of India's cancer burden. The mix also shapes how far the findings travel. They rest most firmly on gynaecological and head and neck queries, and on far fewer lung and colorectal ones.

Seven Measures in One Design

The oncologists assessed seven things: accuracy, clinical safety, guideline concordance, decision impact, physician confidence, efficiency, and agreement between raters.

The authors explain why they wanted all seven in one study. These measures do not necessarily move together. A tool can be accurate and still change nothing in practice, and a tool can be rated highly by its users without being safe. Measuring one and assuming the others gives a partial picture.

To the authors' knowledge, no earlier study of an AI tool in oncology had assessed all seven within a single prospective design in which each clinician serves as his or her own comparison. They also describe the work as the first formal assessment of interrater reliability for an AI oncology evaluation instrument in the Indian context.

Accuracy, safety and guideline concordance were the main endpoints. The paper sets out the statistical methods in full, including the measure of agreement between raters, which was chosen before the study began.

PRECISION-AI Results: Accuracy, Safety and Guideline Concordance

Clinical Accuracy

The oncologists rated 82.7% of answers as fully correct. The paper also reports a composite accuracy rate of 100%. A composite rate counts answers in every category the study's rubric accepts as correct, and the fully correct rate counts only the top category.

Read together, the two figures say that every answer fell within the acceptable range, and that about one answer in six stopped short of the top rating. Both numbers belong in the summary. The fully correct rate is the stricter test, and it is the one to carry in mind when reading any single answer.

Clinical Safety

Every one of the 110 answers was rated clinically safe, a rate of 100%. Safety ratings were also uniform across the three evaluators.

For a tool used close to patient care, this is the result that carries the most weight. An incomplete answer costs the clinician time. An unsafe one could cost a patient far more.

Guideline Concordance

The oncologists rated 94.5% of answers as fully concordant with guideline recommendations, and the composite concordance rate was 100%. Ratings for this measure were uniform across the evaluators as well.

Influence on the Clinical Decision

The oncologists reported that the platform's answer influenced their decision in 92.7% of queries. Influence here includes reinforcing a plan the oncologist already had, as well as prompting a change. The platform summarises the evidence. The clinician always decides.

Physician Confidence

Physician confidence rose by a mean of 1.69 points on a five-point scale (p < 0.0001).

Time to a Clinical Recommendation

The median time to a clinical recommendation fell from 389.2 seconds to 47.2 seconds, a reduction of 87.9% (p < 0.0001). In clinic terms, that is about six and a half minutes reduced to under one minute, or roughly eight times faster.

Study Limitations and How to Read the Results

The authors describe their own study as exploratory and affected by ceiling effects. Four points help in reading the results.

The reference standard is expert judgement. Every figure above is a rating given by a practising oncologist. Patient outcomes such as survival and toxicity were outside the scope of the study. What it shows is that experienced clinicians judged the answers to be correct, safe and in line with guidelines.

Ceiling effects limit what the ratings can separate. When nearly every answer scores at the top of a scale, a study has little room to tell good performance from excellent performance. Harder questions would test the platform more sharply. On agreement, the three oncologists' accuracy ratings were consistent enough for the authors to describe reliability as good (Gwet's AC1 of 0.707).

The sample is modest. The analysis covers 110 queries rated by three oncologists at three centres. More than three quarters of the queries concerned cervical, uterine, head and neck or oral cancers. Breast, lung and colorectal cancers together contributed 25 queries. Confidence in the findings should follow the same proportions.

Earlier research offers context only. The paper cites a 2025 meta-analysis of 56 studies, which reported a pooled accuracy of 76.2% for large language models answering oncology questions across 15 cancer types. Those studies used different tools, questions and scoring methods, so the figure gives context and is not a head-to-head result.

The authors state that deidentified rating data, the rating codebook, the analysis script and the study protocol will be available to qualified researchers on reasonable request. Readers who want every secondary measure, including citation accuracy, will find them in the published abstract.

What the Findings Mean for Oncologists in Practice

For a clinician deciding whether a tool like this deserves a place in the working day, the study supports a measured position.

The most practical finding concerns time. The median time to reach a clinical recommendation fell from about six and a half minutes to under one minute. Across a clinic list, minutes of that kind add up to time that can go back to the patient.

The safety and concordance results suggest the answers are a sound starting point for checking the evidence. About one answer in six was rated short of fully correct, which is the reason the citations are there to be opened and the answer is there to be weighed against the patient in the room.

Comorbidities, performance status, drug access, cost and patient preference sit outside any evidence summary.

Used that way, an evidence summary works like a well-prepared registrar's note. It saves the consultant time, and the consultant still makes the call.

What Comes Next: A Larger Stage II Study

The authors conclude that the results support progression to a larger Stage II study. A larger study can answer what this first stage leaves open: whether the ratings hold with more evaluators and more centres, how the platform performs on cancers that were thinly represented here, and whether harder questions move the scores away from the ceiling.

Try It on Your Own Questions

The most direct test of any evidence tool is a question from your own clinic. OneRx is free for doctors and available on Android and iOS. Get the OneRx app and check its answers against the sources it cites. Hospital teams can read more on the OneRx for hospitals page.

The full paper is available from the journal at https://doi.org/10.1016/j.ijmedinf.2026.106739 .

Frequently Asked Questions

What is the PRECISION-AI study?

PRECISION-AI is a prospective, multicentre study in which three oncologists at three tertiary oncology centres in India evaluated OneRx, a medical AI platform for evidence summary, on 110 clinical queries across five cancer types. It was published in the International Journal of Medical Informatics in 2026.

Was the platform 100% accurate?

On the strictest measure, the oncologists rated 82.7% of answers as fully correct. On the composite measure, which includes every rating category the study accepts as correct, the rate was 100%.

Does the study show that medical AI improves patient outcomes?

Patient outcomes were outside the scope of this study. It measured how practising oncologists rated the platform's answers, along with their confidence and the time they needed. The authors describe the work as exploratory and support a larger Stage II study.

Which guidelines does the platform draw on?

According to the paper, its knowledge base is derived from current NCCN, ESMO, ASCO and ICMR clinical practice guidelines.

Does the platform make treatment decisions?

It does not. The platform summarises published evidence and shows its sources. The clinician always decides.

Who funded the study?

The paper states that there was no funding source and that the authors declared no competing interests.