⚛️ quantum physics

A quantum generative model for in silico clinical trials using scarce training datasets

This paper proposes and validates a quantum generative model pipeline that effectively synthesizes high-fidelity *in silico* patients from scarce clinical datasets, demonstrating superior generalization and expressivity compared to classical baselines using real-world Myelodysplastic Syndrome data on an IBM quantum computer.

Original authors: Olatz Sanz Larrarte, Reza Dastbasteh, Roberto Sanchez-Navarro, Maria Diez-Campelo, Felipe Prosper, Ana Alfonso-Pierola, Mikel Hernaez, Sara Capponi, Pedro Crespo Bofill, Josu Etxezarreta Martinez

Published 2026-08-31

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Olatz Sanz Larrarte, Reza Dastbasteh, Roberto Sanchez-Navarro, Maria Diez-Campelo, Felipe Prosper, Ana Alfonso-Pierola, Mikel Hernaez, Sara Capponi, Pedro Crespo Bofill, Josu Etxezarreta Martinez

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Testing new medicines is a slow, expensive, and often heartbreaking process. Before a treatment can reach the people who need it, researchers must prove it is safe and effective through clinical trials involving real human volunteers. For common illnesses, gathering enough participants is manageable, but for rare or highly complex diseases, finding a large group of patients with the exact same condition is nearly impossible. This scarcity makes it difficult to run the large studies needed to find cures. To solve this, scientists have begun developing "in silico" methods, which use computers to create virtual patients. These digital twins are not real people, but they are built to mimic the statistical patterns of real populations, allowing researchers to run simulations that would be too costly or unethical to perform on living humans. The goal is to generate synthetic data that is so realistic it can help predict how a real group of patients might respond to a new therapy.

However, creating these virtual patients is surprisingly difficult when the real data is scarce. Traditional computer programs used to generate this data usually require massive libraries of patient records to learn the patterns of disease. When researchers only have a few hundred records, as is common with rare conditions, these standard programs often fail, producing fake data that looks nothing like the real thing or simply copying the few examples they were given without understanding the underlying rules. A team of researchers in Spain and the United States has now proposed a different approach, one that borrows from the emerging field of quantum computing. They suggest that the unique way quantum machines handle probability might allow them to learn complex patterns from very small datasets, creating high-quality virtual patients where older methods struggle.

The researchers focused their work on myelodysplastic syndromes, a group of rare blood disorders where patients have a wide variety of genetic mutations and clinical outcomes. They gathered data from a specific clinical trial involving patients with a particular subtype of the disease. The data they had was uneven; they possessed basic information like age, sex, and blood counts for many patients, but detailed information about treatment outcomes and survival was available for far fewer. This "asymmetric" situation, where some details are plentiful and others are missing, is a common hurdle in rare disease research. To tackle this, the team built a new pipeline that first combined these two different sets of information into a single, unified picture of the patient population. They then translated this picture into a format that a quantum computer could understand, using a mathematical structure that compresses complex relationships into a manageable form without losing the essential details.

This compressed representation was then mapped onto a quantum circuit, a specific arrangement of operations on a quantum processor. The researchers ran this process on a real quantum computer located in the United States, specifically a device with 156 quantum bits. The machine generated thousands of synthetic patient records. To see if these digital patients were any good, the team compared them against the real data and against synthetic data created by three well-known classical computer models. They tested the models using different amounts of training data, ranging from just 100 patients up to 400, to see how well each method held up when information was limited. They measured whether the synthetic data matched the statistical distribution of the real patients and whether a computer program could tell the difference between the real and fake records.

The results showed a clear advantage for the quantum approach. When the training data was scarce, the quantum model produced synthetic patients that were much closer to the real population than the classical models. The virtual patients generated by the quantum computer were harder for a classifier to distinguish from real people, indicating that the model had learned the true complexity of the disease rather than just memorizing the few examples it was given. Furthermore, the quantum model was better at reproducing the full range of variations found in the real data, suggesting it was less likely to get stuck repeating the same few patterns. While the classical models performed adequately with larger datasets, they faltered as the number of available patients decreased, whereas the quantum method remained robust. The study suggests that this hybrid approach, which combines classical statistical methods with quantum processing, offers a promising way to generate high-fidelity virtual patients for rare diseases, potentially helping to optimize future clinical trials when real-world data is difficult to obtain.

Technical Summary: A Quantum Generative Model for In Silico Clinical Trials Using Scarce Training Datasets

Problem Statement
In silico methods are increasingly recognized as a strategy to complement clinical trials, particularly for rare or highly heterogeneous diseases like Myelodysplastic Syndromes (MDS), where traditional cohort studies are costly and data is scarce. A critical bottleneck in this domain is the generation of high-fidelity synthetic patient data that accurately reflects underlying clinical variability. While classical generative models (e.g., VAEs, GANs) perform well with extensive databases, they often struggle with the limited sample sizes (often a few hundred patients) typical of rare disease research. Furthermore, classical models rely on rigid parametric assumptions that may introduce bias and fail to capture the true complexity of clinical and molecular distributions. Additionally, clinical datasets are often "asymmetric," containing abundant basic variables but scarce prognostically relevant data (e.g., treatment outcomes).

Methodology
The authors propose a hybrid quantum-classical pipeline designed to address data scarcity and high dimensionality in MDS genomics. The methodology proceeds through the following steps:

Data Integration: The study utilizes two complementary databases from the SintraREV clinical trial: $D_1$ (basic clinical/hematological variables: sex, age, hemoglobin, platelets, neutrophils) and $D_2$ (treatment assignment and survival outcome).
Distribution Estimation: To handle the asymmetric and incomplete nature of the data, the target joint distribution $p(x, y)$ is estimated using the maximum entropy principle. This approach provides the least biased distribution compatible with empirical information, avoiding strong assumptions required by classical models.
Tensor Network Representation: The estimated distribution is encoded into a quantum-compatible state using an ancilla qubit. This state is then compressed into a Matrix Product State (MPS). The MPS serves as a compact latent representation, with its expressive capacity controlled by the bond dimension, avoiding the strong parametric assumptions of classical latent spaces.
Quantum Circuit Mapping: The MPS is mapped onto a deep quantum circuit following a specific encoding scheme and executed on the IBM Heron r2 ("ibm basquecountry") 156-qubit superconducting quantum processor.
Error Mitigation: To counteract noise and limited qubit counts, the pipeline incorporates a simple error-detection and post-selection step based on syndrome verification.
Evaluation: The quality of the generated synthetic cohorts is assessed using metrics for generalization (Mean Absolute Error, Jensen–Shannon Divergence) and expressivity (Random Forest classifier accuracy, Area Under the Curve, and Coverage Ratio).

Key Contributions

Hybrid Pipeline for Rare Diseases: The work introduces a specific pipeline that combines classical statistical inference (maximum entropy), tensor network compression (MPS), and quantum circuit generation to synthesize clinical data.
Handling Asymmetric Data: The method effectively integrates databases with varying levels of completeness (basic variables vs. outcome data) without requiring imputation that might introduce bias.
Empirical Validation on Real Hardware: The model is executed on actual IBM quantum hardware, moving beyond theoretical simulations.
Comparative Analysis: The study provides a direct performance comparison between the proposed quantum model and three classical baselines: Tabular VAE (TVAE), Conditional Tabular GAN (CTGAN), and CopulaGAN.

Results
The model was evaluated on an MDS dataset using training subsets ranging from 100 to 400 patients.

Generalization: The quantum framework consistently achieved lower Mean Absolute Error (MAE) and Jensen–Shannon Divergence (JSD) compared to classical baselines across all sample sizes. The quantum model showed improved performance as the training subset size increased, demonstrating robust generalization in scarce-data regimes.
Expressivity: In Random Forest classification tests designed to distinguish real from synthetic data, the quantum model produced synthetic samples that were statistically indistinguishable from real data (classification accuracy near 50%). It also achieved the highest or near-highest coverage ratios, indicating a strong ability to reproduce diverse patterns and reduce the risk of mode collapse.
Confusion Matrices: The confusion matrices for the quantum model remained close to chance levels, whereas classical models exhibited non-trivial biases, further supporting the superior indistinguishability of the quantum-generated data.

Significance and Claims
The paper claims that the proposed hybrid pipeline offers a competitive and flexible framework for generating high-fidelity in silico patients for rare diseases. The authors assert that their quantum generative model surpasses classical generative models in both generalization and expressivity metrics, particularly when training data is scarce.

The work suggests that quantum computers can effectively represent complex probability distributions to generate high-fidelity synthetic patients, potentially aiding in patient stratification and trial optimization for diseases like MDS. The authors note that while finer discretization of clinical variables could improve expressivity, it would come at the cost of increased quantum resource requirements and sensitivity to noise. The study concludes that the current hybrid approach provides a viable path forward for synthetic data generation in the context of current quantum hardware limitations.

More like this