How Otto Speeds Up Personalized Cardiovascular Modeling

NTT Research has developed a way to personalize cardiovascular computer models without repeating hours of simulation for each patient. The method uses a precomputed library of virtual cardiovascular systems, then adjusts the weight given to each simulation when an individual’s clinical measurements become available.

For technologists working on digital health, the approach addresses a practical bottleneck: how to retain a model’s physiological detail while making patient-specific calculations fast enough to support future clinical use. The NTT Research BioMedical Engineering Innovation (MEI) Lab built the method around Otto, a mechanistic model of the heart and circulation.

Otto represents the cardiovascular system as an electrical analogue. Current corresponds to blood flow, voltage to blood pressure, resistance to opposition to flow and capacitance to the ability of a vessel to expand and hold blood. Connected components describe the heart chambers and circulation. Changing a parameter changes how the system behaves.

In the lung circulation, the researchers linked the resistance values for arteries, capillaries and veins in fixed ratios, and did the same for compliance. The available measurements were too limited to estimate each component independently.

The work is part of the lab’s Bio Digital Twin research program. A cardiovascular Bio Digital Twin is a personalized computer model of a person’s heart and circulatory system. Clinical measurements provide observable features of that system, while the model estimates internal properties that are difficult to measure directly.

To estimate those properties, the study’s comparison method used DREAM, a Bayesian sampling algorithm. It repeatedly ran the model with different parameter combinations and checked the outputs against patient measurements. This search had to be performed separately for each patient.

The peer-reviewed study in Frontiers in Physiology moves much of that work into a reusable virtual population. Researchers constructed the library from published data covering 19 pressure and volume measurements, with at least 125 people in the source sample for each measurement. The library describes simulated healthy cardiovascular systems rather than storing individual patient records.

The team began with 23 model parameters and examined which had the greatest influence on the measurements. A Bayesian sampling method generated plausible parameter combinations. The researchers also adjusted the simulations for differences in body weight and heart rate to better represent variation across a healthy population.

The stored simulations give the method its Bayesian prior: a starting set of possibilities for a patient. To personalize it, the researchers reweighted those simulations using the patient’s pressure measurements and stroke volume, which they calculated from cardiac output and heart rate. Simulations that fit the recorded measurements more closely received more weight.

The result still contains many possible parameter combinations. The study reports ranges for the model’s estimates, including 90 percent credible intervals for pressures and volumes. Where patient measurements are missing, the healthy-population data help constrain the estimates. That also introduces a tradeoff: the results can be pulled toward the population average.

In the reported comparison, creating the virtual population took approximately 78 hours. Once the library existed, the patient-specific calculation took less than two seconds. A conventional sampling method required at least 12 hours per patient with the same software and hardware. The speed gain applies to that calculation, not to collecting clinical data or preparing a system for bedside use.

The team tested the approach with records from 17 intensive care patients without noted cardiovascular disease. Those records came from an independent, de-identified dataset and had not been used to construct the virtual population. The small cohort provided a test of whether the stored simulations could be adapted to individual measurements.

Across the group, the mean absolute percentage error for the measurements evaluated was below 3 percent. The study therefore demonstrated that the library could be adapted to fit the available patient data. It did not test whether the models could predict a patient’s response to medication.

For developers, the reusable library changes where the computing effort sits. A substantial initial calculation can serve multiple patients, while the mechanistic model keeps the relationships among pressure, flow and heart function available for inspection. The same architecture could support research involving many simulated patient profiles.

The authors also propose using virtual populations to test drugs and devices in computer simulations. That would require evidence that the models can predict how circulation changes with treatment. Their suggested approach is to represent a treatment through changes to model parameters, then simulate the response. Those predictions would need validation against patient data.

The present study starts with a healthy population. The models do not yet continuously update as a patient’s condition changes, and the results do not establish clinical decision support. Applying the approach to heart failure or other cardiovascular diseases will require populations that represent those conditions and further validation.

NTT Research plans to develop disease-specific virtual populations, improve drug-effect libraries and investigate models that incorporate new measurements over time. Otto’s faster personalization provides a foundation for that work. Its value in cardiovascular care will depend on whether the models can reliably represent illness and treatment as well as the measurements used to build them.