Tecnun's Synthetic Myelodysplastic-Syndrome Patients Were Seven-Bit Samples From IBM's ibm_basquecountry
Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · August 31, 2026.
The SintraREV trial randomized 61 patients. It ran from 15 February 2010 to 21 February 2018 across 22 university hospitals in Spain, France and Germany, and eight years of enrollment for 61 people is what a phase 3 study costs in a rare blood cancer. A preprint posted on 28 August 2026 takes the data from that trial and asks whether a quantum computer can generate synthetic patients good enough to stand alongside them.
Olatz Sanz Larrarte and nine co-authors published "A quantum generative model for in silico clinical trials using scarce training datasets" on arXiv on 28 August 2026, accepted for the proceedings of the CIBB 2026 conference. The author list crosses two worlds. Sanz Larrarte, Reza Dastbasteh, Pedro Crespo Bofill and Josu Etxezarreta Martinez work in the Department of Basic Sciences at Tecnun, the University of Navarra engineering school in San Sebastián. María Díez-Campelo of the hematology department at Hospital Universitario de Salamanca led SintraREV itself. Felipe Prosper and Ana Alfonso-Piérola practice at the Clínica Universidad de Navarra, Mikel Hernaez and Roberto Sanchez-Navarro sit at CIMA and DATAI in Pamplona, and Sara Capponi is at IBM Research in San Jose. The circuits ran on ibm_basquecountry, the 156-qubit Heron r2 processor in Donostia-San Sebastián.
Practical takeaway. The synthetic patients this pipeline produces are seven-bit strings. Sex, age, hemoglobin, platelets, neutrophils, treatment assignment and survival are each reduced to a single yes-or-no value, and the quantum processor samples from a distribution over those 128 possible patients. The fitting step that classical generative models spend hundreds of training epochs on happens on a classical computer before the quantum machine is touched.
What SintraREV cost to produce: 61 patients, 22 hospitals, eight years of enrollment
The source data comes from a real trial with a full published record. Díez-Campelo and colleagues reported SintraREV in The Lancet Haematology in September 2024, volume 11, issue 9, pages e659 to e670. It tested whether two years of low-dose lenalidomide, 5 mg daily in 28-day cycles, could delay the point at which patients with low-risk or intermediate-1-risk del(5q) myelodysplastic syndromes and anemia become dependent on blood transfusions. Patients were assigned 2 to 1 by telephone randomization, 40 to lenalidomide and 21 to placebo, and the primary endpoint was time to transfusion dependency judged by blinded independent central review. The median age was 72.2 years and 50 of the 61 participants were women. The trial is registered as NCT01243476.
Those numbers explain why the quantum paper exists. Twenty-two academic hospitals across three countries spent eight years to enroll 61 people, in a disease that is uncommon and clinically varied enough that no single site sees many eligible patients. Statistical power in that setting is a physical constraint on how many patients medicine can find, and the authors are direct about the consequence: the high failure rate of phase 3 trials in myelodysplastic syndromes, they write, "underscores the urgent need for improved patient stratification and trial optimization." In silico methods, which build synthetic cohorts to supplement recruited ones, are their proposed answer. The authors cite the United States FDA Modernization Act 2.0 as evidence that regulators are moving, and it is worth being precise about what that statute did. It authorizes certain alternatives to animal testing for the exemption a sponsor needs before investigating a drug's safety and effectiveness, and the agency counts in silico computational models among those alternatives. It says nothing about synthetic participants standing in for recruited ones in a clinical trial, which is a separate question that no instrument has settled.
How the pipeline works: maximum entropy, a matrix product state, and an [8,7,2] parity check
The design starts from a problem every rare-disease researcher will recognize. Basic clinical values are collected on everyone, while the variables that actually carry prognosis are often missing. The paper handles this by splitting the data into two asymmetric databases. D1 holds sex, age, hemoglobin, platelets and neutrophils, written as a five-bit vector. D2 adds treatment assignment to lenalidomide and survival outcome, written as a two-bit vector. Seven variables in total, each of them binary.
The joint distribution over those seven bits is estimated classically using the maximum entropy principle, which the authors choose because it makes the fewest assumptions compatible with what the two databases contain. That distribution is then loaded into an ancilla qubit, compressed into a matrix product state whose bond dimension controls how much correlation survives, and mapped onto a deep quantum circuit following the encoding scheme Shi-Ju Ran published in Physical Review A in 2020. Because the hardware is noisy, samples are filtered by a post-selection step: the distribution is encoded with an [8,7,2] classical parity-check code, and any measured string failing its syndrome check is discarded.
Evaluation is where the paper is at its most careful. Models were trained on subsamples of 100, 150, 200, 300 and 400 patients to see how each behaves as data grows scarce. Three classical baselines, a tabular variational autoencoder, a conditional tabular GAN and CopulaGAN, were trained for 400 epochs on the default SDV architecture. Generalization was scored by mean absolute error and Jensen-Shannon divergence against the target distribution. Expressivity was scored by training a random forest with 100 trees to tell real records from synthetic ones on a balanced set of 1,000 samples with a 70/30 split, where an accuracy near 50 percent means the classifier is guessing, plus a coverage ratio measuring how often the model reproduces target patterns it never saw in training. The quantum pipeline achieves lower mean absolute error and lower Jensen-Shannon divergence than the baselines at every sample size, and the authors qualify the rest carefully: some classical models "remain competitive in specific settings," and the quantum coverage ratio is described as "the highest or near-highest" rather than the highest throughout.
Quantum pillar: computing. Technology readiness: TRL 3 of 9. The authors call this a proof-of-concept, and that is the right word: seven binary variables from one completed trial were sampled on a research processor, which sits far below any prospective study, any regulatory submission and any use in a live clinic.
Where the quantum computer enters: sampling rather than learning
One sentence in section 2.1 changes how the headline claim should be read. "In this work, the model is trained classically using the maximum entropy principle," the authors write, and they present the lightness of that step as an advantage over "the iterative sample-based optimization commonly required by other generative approaches." The learning happens on an ordinary computer. What ibm_basquecountry does is draw samples from the distribution once it has been compressed and encoded.
That is a real and defensible design, and it sidesteps the trainability problems the authors list for quantum neural networks, quantum variational autoencoders and quantum GANs, all of which get worse when training data is scarce. It also narrows what the result demonstrates. The claim supported by this experiment is that a maximum-entropy distribution compressed into a tensor network and sampled on quantum hardware produced better synthetic records, on this seven-variable problem, than three off-the-shelf classical generators run at their default settings. The claim that a quantum computer learned the structure of a patient population is a different and larger one, and the paper does not make it.
The margins are plotted rather than tabulated. Figures 2 and 3 carry the mean absolute error, the Jensen-Shannon divergence, the classifier accuracy, the area under the curve and the coverage ratio as curves against sample size, and Figure 4 gives row-normalized confusion matrices in which values near 0.5 mean the classifier is guessing. A reader can see the ordering of the models and read approximate values off the axes. What no reader can do is quote an exact figure, reproduce a number or compute a difference, because no results table appears in the six-page proceedings version. That is ordinary practice at this length, and it is worth remembering when the finding is summarized elsewhere.
What seven binary variables leave out, and what the preprint does not state
Binarization is the sharpest limitation and the authors say so. Hemoglobin, platelets and neutrophils each become a single bit, and the thresholds that decide which bit are not disclosed. Two patients whose values fall on the same side of an undisclosed cutoff are indistinguishable in the model, however far apart they sit clinically, and age is reduced the same way. A synthetic cohort built like this can reproduce the joint pattern of seven coarse indicators, and it cannot support a dose response, a survival curve or a stratification rule that depends on where inside a range a patient actually sits. The team has already tested finer discretization, reporting that it improves clinical expressivity "at the expense of larger quantum resources and greater sensitivity to noise," and that comparison is held back for the full paper.
A second gap is arithmetic. SintraREV randomized 61 patients, while the models here were trained on subsamples of up to 400. The preprint names SintraREV as the data source and does not state how many records D1 and D2 each contain, nor where the additional records come from. The public paper does not resolve it, and the data and code are available "from the corresponding authors upon reasonable request," so the question is answerable only by asking them.
The hardware detail is thin in the same way. The processor is named and the parity-check code is given, and the shot count, the bond dimension actually used, the circuit depth after transpilation and the fraction of samples discarded by post-selection are absent. Post-selection rates matter here more than usual, because they set how much quantum time a synthetic cohort of useful size would cost. This Monitor covered a heart-disease result last week whose quantum circuits never touched a processor at all. The Tecnun work is on the better side of that line, since its samples came off a real machine, and the reporting has not yet caught up with the experiment.
How Quentir Reads It
The interesting part of this paper is not the quantum processor. It is the framing of the clinical problem. Asymmetric databases, where routine bloods exist for thousands of people and outcome data exists for dozens, describe most rare-disease research and a great deal of ordinary hospital data as well. Treating that asymmetry as the modeling target, rather than as a preprocessing nuisance to be imputed away, is a genuinely useful move, and the maximum-entropy formulation is a principled way to make it. A team could adopt that framing tomorrow with no quantum hardware anywhere in the pipeline.
A sponsor or a hospital research office weighing this approach faces a statistical question and a regulatory one at the same time. Three diligence questions are worth putting to any synthetic cohort, and they are this Monitor's own rather than a published checklist: what provenance exists for the records the generator learned from, what fidelity assessment against the real population has been documented, and what evidence exists that the generator did not memorize identifiable patients. On the third, the paper reports no privacy mechanism and runs no membership or attribute-inference attack, so the privacy properties of this pipeline are untested rather than established. The FDA's draft guidance on externally controlled trials addresses patient-level trial and real-world data and does not settle how a synthetic control arm of this kind would be treated. The IBM-Euskadi Quantum Computational Center opened in October 2025 with biomedicine named as one of its four priority areas under the Basque Government's BasQ strategy, and this project, funded through BasQ as QSynthInSilico, is what that priority looks like in its first year.
The milestone worth watching is already announced inside the paper. The authors have run the comparison between the binarized encoding and a discretized one and are holding it for the full version, and its result decides one thing: whether the advantage over the classical baselines survives once each clinical variable carries more than two ranges, given the larger quantum resources and the greater noise sensitivity the authors say finer discretization costs. If it survives, the argument moves from a conference proceeding to something a trial statistician has to engage with. If the noise eats it as soon as the encoding gets finer, this stays an elegant demonstration of what a 156-qubit machine can sample. Either answer is worth having, and the people who can give it are the same ones who spent eight years enrolling 61 patients.
Sources
Primary source: Olatz Sanz Larrarte, Reza Dastbasteh, Roberto Sanchez-Navarro, María Díez-Campelo, Felipe Prosper, Ana Alfonso-Piérola, Mikel Hernaez, Sara Capponi, Pedro Crespo Bofill and Josu Etxezarreta Martinez, "A quantum generative model for in silico clinical trials using scarce training datasets," arXiv:2608.28168v1, posted 28 August 2026 and accepted for the CIBB 2026 proceedings, from which the two asymmetric databases D1 and D2 and their seven binary variables, the maximum-entropy estimation of the joint distribution, the ancilla qubit and matrix product state encoding, the [8,7,2] parity-check post-selection, the execution on the 156-qubit Heron r2 processor ibm_basquecountry, the training subsamples of 100 to 400 patients, the TVAE, CTGAN and CopulaGAN baselines at 400 epochs on the default SDV architecture, the random forest test on 1,000 balanced samples with 100 trees and a 70/30 split, the coverage ratio, the quoted sentences on classical training, on chance-level confusion matrices and on finer discretization, the QSynthInSilico funding and the data-availability statement are taken. The trial figures come from María Díez-Campelo and colleagues, "Low dose lenalidomide versus placebo in non-transfusion dependent patients with low risk, del(5q) myelodysplastic syndromes (SintraREV): a randomised, double-blind, phase 3 trial," The Lancet Haematology, volume 11, issue 9, September 2024, pages e659 to e670, and its ClinicalTrials.gov registration NCT01243476. The IBM newsroom release of 14 October 2025 and the Basque Quantum pages supply the inauguration date, the 156-qubit Heron configuration and the biomedicine priority of the IBM-Euskadi Quantum Computational Center. Shi-Ju Ran, Physical Review A volume 101, article 032310, 2020, is the matrix product state encoding scheme the pipeline follows. The scope of the United States Food and Drug Administration's draft guidance on externally controlled trials is taken from that guidance document, and the description of the FDA Modernization Act 2.0 as authorizing certain alternatives to animal testing for the exemption needed to investigate a drug's safety and effectiveness, with in silico computational models among the new approach methodologies, is taken verbatim from the agency's own memorandum of understanding 225-25-012 with the National Institutes of Health. The observation that the training subsamples exceed the trial's 61 randomized patients while the preprint states no size for D1 or D2, the reading of binarization as the limit on dose response and stratification, the note on missing shot counts, bond dimension and post-selection rates, and the three diligence questions put to a synthetic cohort are this Monitor's own and are not attributed to the authors or to any regulator.
- "A quantum generative model for in silico clinical trials using scarce training datasets"
- Díez-Campelo and colleagues reported SintraREV in The Lancet Haematology
- NCT01243476
- FDA's draft guidance on externally controlled trials
- IBM-Euskadi Quantum Computational Center opened in October 2025
- BasQ strategy