In the second part of her interview with Healthcare Today, Amanda Randles — director of the Duke Center for Computational and Digital Health Innovation — discusses data quality, model validation challenges, and why computational medicine will transform everyday healthcare within a decade.
Balancing Data Quality and Accessibility
Collecting health data at scale sounds impressive. Collecting the right data, however, is far more difficult. Randles acknowledges this tension directly.
“We are attempting to identify exactly when high-resolution data is essential and when a more streamlined approach will suffice,” she explains.
Her team therefore pursues two parallel paths. First, they use standard commercial wearables. This approach maximises accessibility and ensures findings translate broadly across patient populations. Second, for specific clinical studies, they partner with research labs to develop custom wearables. These devices capture the precise physiological resolution their computational models require.
Why Off-the-Shelf Wearables Fall Short
Consumer wearable devices rarely measure what researchers actually need. Most devices do not directly measure cardiac output or stroke volume. Instead, they estimate these figures — along with blood pressure — from indirect signals. Consequently, they deliver proxies rather than direct physiological readings.
This matters enormously. The reliability of a proxy measurement can shift significantly based on a patient’s body type or skin tone. Randles and her team therefore focus on a critical question: when can clinicians trust these proxies, and in which clinical scenarios must they exercise caution?
Validating a Unique Patient Model
Moving Beyond Population-Level Metrics
Traditional medicine asks whether a patient has crossed a universal threshold. Randles’ approach asks something more nuanced: has this specific patient deviated from their own baseline?
“If a patient’s metrics drop by 15%, the starting point is less important than the drop itself,” she notes.
That shift in framing is powerful for two reasons. First, it makes the model sensitive to meaningful individual changes rather than population averages. Second, it accounts for consistent device bias. If a sensor carries a slight inaccuracy, that bias stays constant over time. Focusing on relative change effectively neutralises it.
Starting With High-Need Patient Groups
For initial medical applications, Randles targets patients who already require clinical intervention — particularly those with heart failure. These patients routinely undergo CT scans or right-heart catheterisations. This gives the team access to invasive, high-fidelity measurements like cardiac output. They then use that data to calibrate a digital model precisely to the individual, rather than relying on population-level estimates.
Addressing Longitudinal Monitoring Challenges
Why Long-Term Validation Is Hard
Historically, measuring 3D blood flow continuously over long periods has been impossible. To overcome this, the team uses multiple validation strategies. They compare models against data from implantable sensors and discrete daily measurements drawn from a wide range of patients. Furthermore, they ask patients to undergo Doppler ultrasound of the carotid artery. The resulting velocity waveforms allow the team to verify that fluid dynamic simulations match real-world observations — both at rest and during exercise.
The Role of Anatomy in Model Accuracy
Individual anatomy plays a decisive role in how blood flows. A specific heart rate, for instance, affects flow differently depending on the unique curves and bends within a patient’s coronary arteries. Therefore, the team validates models across a diverse range of 3D anatomies — not just one arterial shape. Understanding how geometry influences simulation results ensures that the model stays robust across every patient it encounters.
Ethics, Accountability, and Clinical Decision Aids
Supporting Doctors, Not Replacing Them
As digital twin models move toward clinical adoption, questions about ethics and accountability naturally arise. Randles draws a firm distinction. “We are not attempting to replace doctors,” she states. “We are creating clinical decision aids.”
The goal is to give clinicians more comprehensive information and present it in a genuinely actionable form. A doctor cannot quickly process a petabyte of heart rate data spanning two decades. Accordingly, the team conducts user studies with clinicians to determine how best to visualise data — ensuring that physicians assess information accurately and at speed.
How the Monitoring System Works in Practice
In real-world use, the system functions as an intelligent monitoring layer. It flags anomalies as a patient goes about daily life and alerts the treating physician ahead of any crisis. In many heart failure cases, a clinician can then intervene remotely — adjusting a medication dosage or prescribing a statin — without requiring a hospital visit. This model improves quality of care while keeping the doctor firmly in control.
Ultimately, the clinician makes every final decision. The digital twin simply provides a more complete picture of the patient’s health. That distinction, Randles emphasises, sits at the very core of her team’s work.
