AI in Health: From Diagnosis to Drug Discovery

In healthcare the distance between a claim and a deployment is wider than in any other sector. High accuracy on a published dataset guarantees nothing about how the same system behaves on images from your scanners, under your protocol, with your patient population. The question for a manager is not how accurate the model is; it is what it does on our data and inside our workflow.

Two families of use case have come close to practical maturity, and their logic is the same: shortening the initial screening stage. In imaging, the system flags suspicious cases for earlier review and reorders the queue. In drug discovery, a model prioritises candidate compounds for laboratory testing among a very large set. In neither case does the final decision move; what moves is the order and the speed of reaching it.

Claims to take seriously and claims to discount

• Take seriously: Results measured on the data of your own centre, with your protocol and your equipment.
• Read with caution: Accuracy reported on a public dataset, which is usually cleaner and less varied than reality.
• Reject: The promise of replacing clinical judgement; these systems prioritise, they do not diagnose.

Why prioritisation is worth more than diagnosis

In most centres the real bottleneck is not the accuracy of the specialist; it is time. A queue reviewed in order of arrival can leave a high-risk case waiting for days. A system that reorders that queue by probability of risk, without deciding anything, shortens time to action. This is where measurable value and manageable risk meet, provided no case is dropped and every one still reaches a specialist.

The measure that matters in healthcare

Overall accuracy is a misleading measure in clinical use. What matters is the distribution of errors: the cost of a false negative, a case the system calls safe when it is not, is not comparable to the cost of a false positive. A system tuned for screening should lean toward high sensitivity and accept extra alerts. Setting that balance is a clinical decision rather than a technical one, and it belongs in writing before deployment.

Why it matters for Iranian SMEs

Iranian clinics and laboratories hold a large volume of clinical data generated by the patient population of this country, precisely what an imported solution lacks. For health-sector companies that is the chance to build a solution validated on local data. The starting point needs no large budget, but it has two serious preconditions: a clear framework for protecting patient data, and a clinician involved from day one rather than at the end for sign-off.

A 90-day map for starting in healthcare

• Days 1 to 30: Choose one frequent, low-risk process, measure its current time to result, and put the data-protection framework in writing.
• Days 30 to 60: Evaluate the system on the data of your own centre and review the error distribution with a clinician, not just overall accuracy.
• Days 60 to 90: Run it in an assistive, parallel mode; if time to action fell and costly errors did not rise, widen the scope step by step.

Recurring mistakes

• Relying on vendor-stated accuracy without testing on the data of your own centre.
• Judging performance by overall accuracy and ignoring the asymmetric cost of errors.
• Bringing the clinician in at the end for approval instead of from the design stage.
• Starting with a high-stakes use case rather than a process where errors are recoverable.

Three moves for this quarter

• Choose one frequent, low-risk process and record its current time to result.
• Involve a clinician from day one in defining the use case and the error balance.
• Run the evaluation on your own data and set the result beside the vendor figure.

Frequently asked questions

• Do these systems replace physicians?
  In the mature use cases of today, no. They order and prioritise the queue; the decision and the responsibility stay with the specialist.
• How do we test the claim of a vendor?
  By running it in parallel on your own historical data and comparing the error distribution, rather than trusting an accuracy figure from a public dataset.
• How do we protect patient data?
  Through data minimisation, removing identifiers before processing, defined access levels, and a written retention and deletion policy.

Takeaway

In healthcare the value of AI today lies not in moving the decision but in shortening the path to it. That is also the right measure for a manager: did time to action fall, and did costly errors stay flat? Any claim that cannot be answered with those two numbers on your own data remains, until further notice, a claim.

Glossary

• Screening: The initial stage of filtering cases or options for closer examination.
• Diagnostic support: A system that flags suspicious cases for specialist review without issuing a diagnosis itself.
• False negative: A case the system calls safe when it is in fact high-risk; the costliest error in screening.
• Sensitivity: The ability of a system to find genuinely high-risk cases, even at the price of extra alerts.
• Parallel run: Operating a system alongside current practice to compare results before relying on it.