Pharmacist decision support · Founded 2025

AI flags the interaction, the pharmacist makes the call.

Drug Interaction Check is a decision-support concept for screening drug interactions in patients on heart and diabetes medicines. It cross-checks a validated database against AI chatbot output and shows the pharmacist where they disagree, so a qualified person always decides.

See the solutionView the results
150cardio-diabetic drug pairs benchmarked
6resources compared: 3 databases, 3 AI chatbots
κ = 0.51mean pairwise agreement (moderate)
Benchmark snapshot At a glance
LexicompReference standard · interacting pairs
105 / 150
ChatGPTSensitivity / specificity vs reference
0.98 / 0.10
Drugs.comSensitivity / specificity vs reference
0.84 / 0.38
Why it matters: Chatbots caught almost every interaction but also flagged pairs the reference did not, which is why a pharmacist stays in the loop.
The problem

Interaction checkers do not agree with each other.

Patients with hypertension and diabetes often take several medicines at once. The tools used to screen those combinations give different answers.

Severity is inconsistent

Across six resources, agreement on interaction severity was only fair (Fleiss' kappa 0.28), and slight for major interactions (0.15).

Chatbots over-alert

AI chatbots reached sensitivity up to 0.98 but specificity as low as 0.02 against the reference database, so many alerts may be unnecessary.

Databases can miss detail

One free database covered clinical management in just 2.46% of entries and risk factors in 5.74%, leaving pharmacists to fill the gaps.

Our solution

A hybrid human and AI screening workflow.

The study recommends validated databases for the first check, chatbots for secondary screening, and qualified review of every AI output. Drug Interaction Check is designed around exactly that order.

First check

Validated database

100%scope of the reference database in the benchmark
  • Every pair starts with a validated interaction source
  • Severity taken from the reference, not the chatbot
  • Clear record of which source said what
Second look

AI cross-check

0.96 to 0.98sensitivity of the two most sensitive chatbots
  • Chatbots used to catch possible misses
  • Answers shown with full completeness on effects and mechanism
  • Never used as the primary source
Final decision

Pharmacist review

0.28Fleiss' kappa for severity: why a human decides
  • Disagreements highlighted for review
  • Pharmacist confirms severity and management
  • Tool supports, never replaces, clinical judgement
How it works

How it works

Four steps, modelled on the benchmark method.

  1. 1

    Enter the drug pair

    The pharmacist enters the combination, for example an antihypertensive with an antidiabetic agent.

    antihypertensiveantidiabetic
  2. 2

    Database check

    The pair is checked against a validated interaction database first, as the study recommends.

    reference first
  3. 3

    AI screening

    A standardised prompt is sent to AI chatbots in a fresh session, mirroring how the benchmark queried them.

    standard promptnew session
  4. 4

    Flag and review

    Where sources disagree on presence or severity, the pair is flagged for pharmacist review before any action.

    disagreement flaghuman review
Results

What the benchmark showed

A comparative evaluation of 150 literature-derived drug pairs across three databases and three free AI chatbots, with Lexicomp as the reference standard. This is a tool comparison, not a clinical trial.

150drug pairs with human evidence, finalised by an expert panel
0.28overall Fleiss' kappa for severity (fair)
0.51mean Cohen's kappa, pairwise (moderate)
100%completeness for all three chatbots
0.02lowest chatbot specificity vs reference
Performance against the reference database (Lexicomp)
ResourceScopeCompletenessSensitivitySpecificity
Lexicomp100%85.2%ReferenceReference
Drugs.com94%90%0.840.38
DrugBank97.33%61.6%0.700.48
ChatGPT100%100%0.980.10
Copilot99.33%100%0.960.22
Gemini100%100%0.870.02

Source: Alam et al., International Journal of Applied Pharmaceutics 17(5), 253-262 (2025). doi:10.22159/ijap.2025v17i5.54398

Who it's for

Who it is for

People who check medicine combinations every day.

Community and clinical pharmacists

  • Faster second look at complex lists
  • Disagreements surfaced, not hidden
  • Final call stays with the pharmacist

Pharmacy educators

  • Real examples of where AI and databases differ
  • Teach critical appraisal of AI output
  • Benchmark data for class discussion

Health systems and app builders

  • Evidence on chatbot over-alerting
  • A hybrid workflow to adopt
  • Clear limits on AI use
Where we are

What we do not claim

Honest limits from the underlying study.

Not a replacement for a pharmacist

AI output can include unsupported claims and varies with prompt wording. Every result needs review by a qualified clinician or pharmacist.

One reference standard

Accuracy was measured against Lexicomp, so the benchmark inherits that database's own classifications.

Research stage

The benchmark covers 150 cardio-diabetic pairs and free chatbot versions. It has not been tested in patient care.

The team

The team

Founded in 2025 by the research team behind the cardio-diabetic drug interaction benchmark.

AA

Aftab Alam

Co-author of the Drug Interaction Check research.

AS

Anukriti Saran

Co-author of the Drug Interaction Check research.

RJ

Radhika Joshi

Co-author of the Drug Interaction Check research.

SS

Sathvik Belagodu Sridhar

Co-author of the Drug Interaction Check research.

SS

Swapnil Sharma

Co-author of the Drug Interaction Check research.

SP

Sarvesh Paliwal

Co-author of the Drug Interaction Check research.

Frequently asked questions

Does this replace a pharmacist?

No. It is decision support. The study recommends that AI output is reviewed by qualified clinicians or pharmacists, and the tool is built around that review.

Why not just use an AI chatbot?

In the benchmark, chatbots were highly sensitive but had low specificity, down to 0.02, and inconsistent severity ratings. That risks over-alerting.

Which drugs were studied?

150 pairs, each including at least one antihypertensive or antidiabetic medicine, selected from the literature and confirmed by an expert panel.

How much did the tools agree?

Overall agreement on severity was fair (Fleiss' kappa 0.28). Pairwise agreement on whether an interaction exists was moderate (mean Cohen's kappa 0.51).

Is it available now?

Not yet. Drug Interaction Check is a research-stage concept based on a published comparative study.

The research

Peer-reviewed science behind Drug Interaction Check.

Read the full study behind the product.

Read the paper

Alam, A., Saran, A., Joshi, R., Sridhar, S. B., Sharma, S., & Paliwal, S. (2025). From algorithms to evidence: Assessing artificial intelligence chatbots and drug databases for detecting cardio-diabetic drug interactions. International Journal of Applied Pharmaceutics, 17(5), 253–262. https://doi.org/10.22159/ijap.2025v17i5.54398

https://doi.org/10.22159/ijap.2025v17i5.54398