The next medical test could be chosen by an AI that first imagines what the result might say

How ACTMED turns imagined results into a choice of tests

Imagined results become a choice.

ACTMED asks a language model to picture plausible outcomes for each candidate test, then scores how strongly those outcomes would move the disease probability. In a worked chronic kidney disease example, the system compares serum creatinine and sodium. For each candidate test it imagines two plausible results, computes the updated chance of kidney disease after each, and then asks which test would most change that chance if ordered.

A high creatinine value lifts the probability to 0.65, while a normal one pulls it back to 0.22. A low sodium result reaches only 0.45, and a normal sodium reading drops it to 0.18. Because creatinine’s imagined outcomes would shift the probability more, ACTMED selects serum creatinine over sodium.

The team calls this expected information gain, a measure of how much a result would sharpen the diagnosis. The method treats each test as an experiment whose value is not its cost alone but the change in belief it can produce before any sample is drawn. Ruhrberg Estévez and colleagues write that “the framework issues test recommendations rather than fixed decisions, allowing clinicians to override suggestions based on context and re-query with alternative test results.”

Why the quality of the simulator decides how well the system works

The simulator sets the ceiling. ACTMED chooses a test by first imagining what its possible results might be, then asking which result would move the diagnosis most. If those imagined outcomes are wrong, the choice that follows is wrong too.

A weak simulator can fail in two different ways. One open-source model sometimes produced clinically implausible values, such as non-integer numbers for previous pregnancies. Another often returned the same value across repeated samples even when sampling was made more random. In both cases, the system could not estimate how much a new result would change belief.

That failure shows up in accuracy. On hepatitis C, the weaker medical simulator made ACTMED less accurate than a simple classifier that used every available test (0.489 versus 0.511). Ruhrberg Estévez and colleagues suggest this happens because such models struggle to capture patient-specific uncertainty.

A stronger general-purpose simulator did better: its simulated outcomes stayed closer to real distributions, and ACTMED beat random selection, a fixed best-test set, and implicit LLM selection on the broader OSCE-style case set. The team notes that current large language models are not yet ready for direct autonomous diagnosis.

Where clinicians stay in control, and where the method still falls short

Clinicians remain the deciding voice. The system can suggest tests, but a doctor can inspect starting risk estimates, check simulated outcomes for plausibility, reject infeasible tests, and override unsafe choices. Ruhrberg Estévez and colleagues write that “current LLMs are not yet ready for direct autonomous diagnosis”.

The team also notes that “our framework is also currently limited to binary classification. Each task is framed as a single yes or no decision, even though the work includes a broader OSCE style case collection. Future evaluation should still cover co-morbidities and multi-label settings if the method is to reach real wards.

Cost is another limit. Estimating the expected value of one candidate test took about 20 seconds with GPT-4o or GPT-4o-mini, while a static classifier can use all features in a single pass. The authors argue that this latency is acceptable because unnecessary testing and delayed diagnosis carry larger medical and financial costs.

The method also leaves out free-text reports such as imaging or pathology notes. Structured versions of those records could be added later, but until then the system works only with categorical and numerical features.

## A useful assistant, not a replacement

The value of ACTMED is that it turns diagnosis into a sequence of choices instead of a single guess. The system asks what each test would likely reveal, then recommends the one expected to reduce uncertainty most. That makes it different from a black box that simply outputs a label.

Ruhrberg Estévez and colleagues frame ACTMED as issuing test recommendations rather than fixed decisions. They also note that current LLMs are “not yet ready for direct autonomous diagnosis” but may be better suited as decision-support tools. The opportunity is real: fewer unnecessary tests, faster triage in busy clinics, and a record of why each test was chosen.

Doctors can stay in the loop, reviewing test suggestions, interpreting intermediate outputs, and applying clinical judgment throughout. The method depends on simulators that can imagine plausible results, a dependence that weaker models expose when they produce values no patient would have or collapse into one repeated answer. The evaluation included a broader OSCE-style case collection with varied cases, but each task is framed as a binary diagnostic decision.

The framework was applied across distinct diseases rather than within one multi-label setting, and the team points to co-morbidities as future work. Free text from scans and pathology reports is still outside its reach. Even so, the direction points toward medicine in which machines help choose the next step while doctors keep final judgment.