Florida Atlantic's 90.26% Quantum Heart-Disease Result Came From a Simulator: the AI Paper of 21 May 2026 and the FAU Release of 27 August
Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · August 30, 2026.
Florida Atlantic University announced on 27 August 2026 that its engineers had built a quantum machine learning framework for heart disease prediction reaching more than 90 percent accuracy. The figure is precise and it is checkable. It is 90.26 percent, produced by a quantum support vector machine with angle encoding, averaged across five folds of a clinical file holding 918 patients.
The work behind that announcement appeared three months earlier. Muhammad Minoar Hossain, Md. Hasibul Hassan Himal and Arslan Munir published "A Comparative Study of Quantum Feature Maps and Quantum Classifiers for Heart Disease Prediction" in the MDPI journal AI on 21 May 2026, as article 180 of volume 7, under a Creative Commons license that lets anyone read the whole methods section. Section 2.6.2 of that paper states that every quantum experiment ran in a simulation-based environment rather than on a physical quantum processing unit. The university announcement does not carry that sentence, and neither does the coverage that followed it.
That difference decides what the study is evidence for. Simulated qubits establish whether an algorithm has promise in principle; a run on a physical processor establishes whether the machines that exist can deliver it, once noise, limited connectivity and readout error have had their say. Everything else in the paper holds up well under checking. Several of its numbers are more informative than the ones the announcement chose to lead with, and one of them disagrees with the paper's own abstract. The study is a careful piece of comparative work whose careful parts were the first thing lost in transmission.
Practical takeaway. No quantum computer was involved in this result. Nine logical qubits were imitated in software on an ordinary machine, fed nine principal components drawn from 918 public patient rows, and the winning quantum kernel reported a 0.76-percentage-point higher mean accuracy than a classical support vector machine under the paper's own comparison, with no confidence interval or significance test attached to that gap. In their own conclusion the authors write that the findings "should not be interpreted as proof of general superiority over classical ML or as a direct claim of clinical readiness."
What Hossain, Himal and Munir ran: 918 Kaggle rows, nine principal components, nine simulated qubits
The clinical data came from the Heart Failure Prediction Dataset assembled by the Kaggle contributor Fedesoriano in 2021, which the paper's data availability statement gives as its source and which the authors accessed on 17 May 2026. That file holds 918 unique patient rows and was built by combining five older heart disease databases. It is one of the most heavily used benchmarking files in medical machine learning, and that matters for how the accuracy figure should be read.
Preprocessing reduced the clinical variables to nine principal components. Those nine components were mapped onto nine qubits, and the same nine were used across every experiment in the study. The team then built a grid. Five ways of encoding classical numbers into quantum states, being angle encoding, amplitude encoding, basis encoding, Pauli encoding and the ZZ feature map, were crossed with four quantum classifiers, being a quantum support vector machine, a quantum k-nearest neighbor, a quantum random forest and a variational quantum circuit. Every combination was evaluated with five-fold cross-validation, splitting 80:20 inside each fold, on gate-based circuits with a linear entanglement pattern.
Angle encoding with the quantum support vector machine won. Table 8 of the paper gives its fold-by-fold results and their means: accuracy 0.9026, specificity 0.8342, sensitivity 0.9216, precision 0.8739, F1 0.8968, Cohen's kappa 0.7608, and an area under the ROC curve of 0.9300. One number in the abstract does not agree with that table. The abstract reports precision of 88.89 percent, while the five fold precisions in Table 8 average 87.39 percent. The discrepancy sits in the published paper. It went no further, because the announcement and the news items that followed carried no precision figure at all, and a reader quoting the abstract would still be quoting the higher of the two.
Why 90.26 percent sits 0.76 points above a classical support vector machine on the same data
The paper answers the obvious question itself. In Figure 16 the authors set their quantum support vector machine against an ordinary classical one using linear, polynomial and radial basis function kernels under the same evaluation setting. The classical model with an RBF kernel reaches 89.5 percent. The quantum model reaches 90.26 percent. The authors read that gap as evidence that "the quantum feature space provides slightly better class separation than standard classical kernels", and the word slightly is theirs. No confidence interval and no significance test is reported for the difference, so what the paper establishes is a 0.76-percentage-point higher mean accuracy under its own comparison, on 918 rows, across five folds.
They are equally careful with the literature. Table 9 lists earlier heart disease studies claiming 96.4 percent, 97.06 percent and even 99.96 percent accuracy, and the authors decline to treat those as the bar to clear. Most skipped cross-validation, several leaned on hybrid models in which the classical components did the real learning, and the highest figures come from a single train-test split that flatters the model. Their own claim is narrower and better supported.
A second file supplies a further check, and it should be read for what it is. The authors ran the winning model against a heart disease dataset published on Kaggle by R. Mammadov in 2022, 1,000 patient rows with 15 variables, accessed on the same day as the primary file, and obtained 89.69 percent accuracy with an area under the curve of 0.9186. The paper does not state whether the model was retrained on that file or transferred unchanged, and the Kaggle page carries no collection protocol, no site and no clinical provenance. That makes the exercise a secondary robustness check on a second tabular file rather than external clinical validation, and the figure lands below the headline number.
Quantum pillar: computing. Technology readiness: TRL 3 of 9. The quantum circuits exist only as software on an ordinary computer that imitates qubits, exercised on a public file of past patient data, which sits several stages before any hospital system, any prospective study and any regulatory review.
What the simulator left out: no transpilation, no gate noise, no readout error
Section 2.6.2 is worth quoting, because it is the part that did not travel. "All quantum experiments in this study are conducted in a simulation-based environment rather than on a physical quantum processing unit," the authors write, and "the qubit counts discussed in this work refer to logical qubits required by the encoding schemes and do not represent hardware-mapped physical qubits." Because the circuits were never compiled down to a specific machine, the paper says that physical qubit allocation, coupling map constraints, routing overhead, calibration variability, gate noise and readout error "were not explicitly incorporated into the benchmarking pipeline". Its statement about hardware efficiency is therefore, in the authors' words, "a qualitative consideration rather than as a backend measured result".
That is an honest engineering position, and the authors repeat it in their future work, where they plan to validate the method on noisy quantum hardware because the present results "do not capture practical issues such as noise, qubit topology, routing". The distance between a nine-qubit noiseless simulation and a nine-qubit run on a real superconducting processor is not a rounding error. Current devices lose fidelity with every two-qubit gate, and an entangling layer across nine qubits followed by a kernel estimate is the kind of circuit where that loss tends to show. How much it would cost this particular model is unknown, because no backend, transpilation, shot count or noise experiment appears in the paper. The 0.76-percentage-point margin over the classical kernel could survive a run on hardware, shrink, or reverse, and only that run will say which.
Hardware work from the same week shows the difference. On 27 August this Monitor covered A*STAR's molecular docking run on six qubits of IBM's ibm_kingston, where the calculation touched a real processor and the team had to report what the noise did to it. Both kinds of study are legitimate. They answer different questions, and only one of them says anything about a machine a hospital could buy time on today.
Why FAU's 27 August release leaves a clinician unable to find the paper
The university announcement describes the venue as "the MDPI AI Journal (impact factor of 6.5)". It gives no volume, no article number, no DOI and no publication date. A clinician or a hospital informatics lead who wanted to read the methods would have to go looking. The impact factor is a journal-level citation average and says nothing about this particular study, and it is the only quality signal the announcement offers.
The announcement also credits the work to a group "led by Arslan Munir". Munir is the corresponding author and directs FAU's Intelligent Systems, Computer Architecture, Analytics and Security laboratory. The first author is Muhammad Minoar Hossain, in the same FAU department, and the second author is Md. Hasibul Hassan Himal at Bangladesh University in Dhaka. Neither of them appears in it.
From there the figures moved unchanged. The Quantum Insider carried the story on 27 August, News-Medical on 29 August, and a French rendering on the clinical site ma-clinique.fr reached francophone readers the same day. Every version repeats 90.26 percent accuracy, 92.16 percent sensitivity, 83.42 percent specificity and an AUC of 0.93. None of them uses the word simulator.
How Quentir Reads It
Read for what it is, this is a competent comparative study, and it does the thing the field mostly skips. It puts five encoding schemes and four classifiers on the same footing, uses cross-validation throughout, runs a second file as a robustness check, and reports a classical support vector machine evaluated, in its own words, under the same setting. What that comparison does not document is a matched hyperparameter-search budget on either side, so it establishes a like-for-like run rather than a tuned contest. The finding that survives all of that is modest and worth having. On a small tabular clinical problem, a quantum kernel evaluated in simulation performs about as well as a good classical one.
For a hospital the operative figures are the two the headline hides. A specificity of 83.42 percent means roughly one in six patients without heart disease is classified as having it, which in a screening context is a large volume of downstream imaging and cardiology referrals. Sensitivity of 92.16 percent means about one patient in thirteen who does have the disease is missed. Those are the numbers a clinical governance committee weighs, and they are ordinary for this problem rather than remarkable.
The file itself is also not a hospital population. It was assembled from five older collections for benchmarking, and a model fitted to it has never met a live clinic, a different hospital site, a different exercise-test and measurement workflow, a different coding practice or a patient group it was not built from. That gap is the usual reason retrospective accuracy figures shrink on first contact with a real service, and nothing about a quantum kernel changes it.
The authors themselves supply the sentence a procurement conversation should start from, and it is the one the announcement chain dropped: the findings "should not be interpreted as proof of general superiority over classical ML or as a direct claim of clinical readiness". The next milestone that would move this work forward is already stated in the paper, which is the same comparison run on noisy hardware. Whenever Munir's group publishes that, it will be the version worth reading, because it is the first one in which the word quantum refers to a machine.
Sources
Primary source: Muhammad Minoar Hossain, Md. Hasibul Hassan Himal and Arslan Munir, "A Comparative Study of Quantum Feature Maps and Quantum Classifiers for Heart Disease Prediction," AI 2026, volume 7, article 180, published 21 May 2026 (received 25 February 2026, accepted 18 May 2026), from which the 90.26 percent accuracy, the fold-by-fold means in Table 8, the classical RBF-kernel comparison of 89.5 percent in Figure 16, the comparison with prior work in Table 9, the second-file result of 89.69 percent accuracy and 0.9186 AUC on the 1,000-row Kaggle dataset published by R. Mammadov in 2022 that the paper cites as reference 34, the nine principal components mapped to nine qubits, the five feature maps and four classifiers, the quoted passages of Section 2.6.2 on simulation-based execution, and the conclusion and future-work sentences are taken. The Florida Atlantic University newsdesk announcement of 27 August 2026 supplies the public claim of more than 90 percent accuracy, the description of the venue as the MDPI AI Journal with an impact factor of 6.5, the attribution of the work to a group led by Munir, and the quotations from Munir and from dean Stella Batalama. The Kaggle Heart Failure Prediction Dataset published by Fedesoriano in 2021 is the source file the paper names in its data availability statement, accessed by the authors on 17 May 2026. The Quantum Insider of 27 August 2026, News-Medical of 29 August 2026 and ma-clinique.fr of 29 August 2026 are cited for how the figures travelled onward. The observation that the abstract reports 88.89 percent precision while the fold means in Table 8 average 87.39 percent is this Monitor’s arithmetic on the published table. The reading of specificity as a referral burden, the description of the second Kaggle file as a robustness check rather than external clinical validation, the assessment of what a retrospective benchmarking file does not represent, and the comparison with hardware-based work are this Monitor’s assessment and are not attributed to the authors or to the university.