Support, not a verdict
Diagnostic AI is usually best understood as decision support. The software looks at a scan, a tissue image or a set of test results and produces something for a clinician to consider: a flag, a ranking, a highlighted area or an estimated probability. The clinician, who knows the patient, weighs that against everything else. This article looks at that narrower use, how these tools work in general terms, and why their accuracy needs careful reading.
This is general information only. It cannot tell you anything about your own health, and no tool described here should replace a conversation with a qualified professional.
What the tools look at
Most diagnostic tools are built for a specific, narrow input:
- Medical images, such as X-rays, scans or photographs of tissue or skin.
- Signals, such as heart recordings.
- Laboratory results, where a model looks at combinations of values.
- Combined records, bringing together several data sources.
Each tool is typically designed for one task on one kind of data. A program trained to look at one type of image does not generalize into a doctor who understands everything. Outside the situation it was built for, its output may be meaningless.
How they are built
Developers collect many examples where experts have already established the answer, and then train a model to link input to label. For images, this often involves neural networks, which learn to pick up visual patterns from thousands of examples. After training, the model is tested on examples it did not see during training to estimate how it performs.
The key point is that the quality of the examples sets the ceiling. If the labels came from experts who sometimes disagreed, or if certain kinds of patients were rare in the training data, the tool inherits those weaknesses.
Ways it can help
Possible roles include:
- Prioritizing. Sorting a queue so cases that look more urgent are reviewed sooner.
- A second look. Highlighting something a busy reader might overlook.
- Consistency. Software does not get tired, though it can be consistently wrong in the same way.
- Measurements. Automating routine measurements that are tedious to do by hand.
- Access. In places with few specialists, a screening aid might help decide who needs referral, though this depends on the tool being proven in that setting.
Studies have compared some tools with clinicians on particular tasks, with varying results. Performance in a controlled study does not always carry over to everyday practice.
Why accuracy is more complicated than one number
When a tool is described as “95% accurate,” that statement alone says little. Several ideas matter.
Sensitivity and specificity. Sensitivity is how often the tool correctly flags people who do have the condition; specificity is how often it correctly clears people who do not. A tool can do well on one and poorly on the other.
How common the condition is. Here is a hypothetical example, with made-up numbers chosen only to show the arithmetic. Suppose a condition affects 1 in 100 people screened, and a tool catches 90% of true cases and wrongly flags 10% of healthy people. Among 1,000 people, about 10 have the condition and the tool flags 9 of them. Of the 990 without the condition, it wrongly flags 99. So about 108 people are flagged, and only 9 of them truly have it, roughly 1 in 12. The takeaway: when a condition is uncommon, even a seemingly strong tool can produce many false alarms. Real tools and real conditions differ, so these figures illustrate the logic and are not a statement about any actual product.
Test conditions. Results from a carefully selected research dataset may not hold up with different scanners, different hospitals or different patient groups.
Measuring the right thing. Matching an expert’s reading is not the same as improving patient outcomes. A tool can detect more findings without that helping anyone, for example by finding small abnormalities that would never have caused harm.
Where things go wrong
- Bias in the data. If some groups were underrepresented during training, performance may be worse for them, a form of algorithmic bias with real consequences in medicine.
- Shortcuts. A model may rely on incidental clues, such as markings or equipment differences, instead of the condition itself.
- Drift. Changes in equipment, procedures or populations can erode performance over time.
- Overreliance. If staff begin to defer to the software, errors in its output may go unnoticed. The reverse can also happen: staff may ignore a tool that has raised too many false alarms.
- Opacity. Many models cannot easily explain their output. Work on explainable AI, such as highlighting which part of an image influenced a result, is helpful but does not guarantee the explanation reflects what the model actually relied on.
Why clinicians stay in charge
A diagnosis is more than a label. It involves symptoms, history, examination, other test results, a patient’s preferences and a judgment about what to do next. Clinicians carry professional and legal responsibility for those decisions. Human oversight is therefore not a formality: it is the safeguard that catches the tool’s mistakes and fits its output to the individual.
Questions of accountability, such as who is answerable if a tool contributes to a missed or wrong diagnosis, are still being worked through by professional bodies, hospitals and lawmakers, and the answers vary by place. Many jurisdictions treat diagnostic software as a regulated medical device that needs review before use, and the broader landscape of AI regulation continues to evolve.
What patients can reasonably expect
You may never be told that software played a part in reading your scan. It is fair to ask a provider how such tools are used, whether a clinician reviews every result and how your data is handled. If a result worries you, or an app or website gives you a reading you do not understand, take it to a qualified professional. Do not use an online tool to rule a condition in or out on your own.
Diagnostic AI has real promise in narrow tasks, and its value will probably be decided by careful testing, honest reporting of errors and good integration into clinical work, not by bold headlines.