Abstract
Infrared molecular fingerprinting of blood offers a scalable, minimally invasive window into human physiology, but limited follow-up, imbalanced cohorts, and restricted access to diverse phenotypes constrain systematic studies. Here, we introduce a conditional deep generative framework for synthesizing blood-based infrared spectra that preserves individual-level structure while allowing controlled manipulation of demographic and anthropometric covariates. Using 25,308 spectra from 5,863 ostensibly healthy participants in the longitudinal Health4Hungary – Hungary4Health cohort, we train a Conditional Variational Autoencoder and a Conditional Boundary Equilibrium GAN to generate blood-based infrared spectra conditioned on age, sex, and body mass index. We show that the generated spectra closely match held-out real data across multiple levels and faithfully encode demographic and anthropometric information. We further demonstrate two in-silico applications: modeling of individualized healthy aging trajectories that follow cohort-level aging manifolds while retaining subject-specific characteristics, and targeted augmentation of underrepresented body mass index categories. Together, these results establish conditional generative modeling of blood-based infrared spectra as a viable approach for virtual cohort construction, cohort balancing, and controlled in-silico phenotyping, paving the way toward more comprehensive and data-efficient studies in precision health.
Introduction
Infrared Molecular Fingerprinting (IMF) of blood is a powerful tool for monitoring health by capturing the vibrational signatures of molecules in plasma. This provides a holistic snapshot of an individual's biology. However, the utility of IMF is often limited by the logistical challenges of large-scale longitudinal studies and the inherent imbalances in clinical datasets. For example, data for specific age groups or BMI profiles may be underrepresented, hindering systematic analysis of aging and disease progression.
To overcome these limitations, this research introduces a conditional deep generative framework designed to synthesize realistic blood-based infrared spectra. By creating "digital twins" of real samples, researchers can perform controlled in-silico experiments, such as simulating aging or balancing datasets for more robust machine learning models.
Methods and Data
The study utilized 25,308 infrared spectral samples from 5,863 healthy volunteers in the Health4Hungary – Hungary4Health (H4H) cohort. This longitudinal dataset includes participants with multiple visits, providing a rich basis for modeling biological variability.
The framework employs two primary deep learning architectures:
- Conditional Variational Autoencoder (CVAE): Used to encode unique spectral baselines into a latent space, enabling smooth manipulation of variables like age while preserving individual identity.
- Conditional Boundary Equilibrium GAN (CBEGAN): An adversarial model focused on generating high-fidelity spectra from noise, accurately reproducing sharp, condition-specific spectral features.
Both models were conditioned on age, sex, and BMI. This conditioning ensures that the generated data is not just an "average" sample but reflects specific demographic and anthropometric characteristics.
Results
The synthetic spectra exhibited high fidelity to real-world data. The effect size of differences between real and generated spectra remained below 30% of the standard deviation across all wavenumbers, suggesting that deviations are of limited practical significance. Furthermore, the models preserved 24 specific biochemical peak ratios, ensuring that the synthetic data remains biologically interpretable.
In predictive tasks, models trained on synthetic data showed remarkable consistency with real-data baselines. For sex classification, a CBEGAN-trained model achieved an AUC of 0.95, nearly matching the 0.96 AUC of the real-data baseline.
The framework also proved effective for data augmentation. In a BMI classification task where data for certain groups was artificially reduced (dropping accuracy to 38.1%), augmentation with CBEGAN-generated spectra restored accuracy to 55.9%. This demonstrates the potential for generative models to mitigate the impact of imbalanced datasets in medical research.
Discussion and Future Directions
While the current framework conditions on age, sex, and BMI, future iterations could incorporate additional clinical parameters such as lipid panels or glucose levels to further refine the models. Additionally, while this study focused on healthy participants, extending the framework to specific disease states, such as oncology, is a key area for future work.
Ultimately, the transition from conditional modeling to causal modeling will allow for true counterfactual "what-if" simulations. This would enable clinicians to predict how a specific individual's molecular fingerprint might change in response to lifestyle interventions or medical treatments, moving us closer to the goal of personalized precision health.
Summary updated