Patient heal thyself: The promise of physiological avatars in predictive healthcare analytics

Since it was first coined in the 1990s, the term “big data” has been weighed down by expectations of analytical revolutions in science, technology, and industry. Over the last thirty-plus years, information storage capacity has grown, and computing power has increased to create a big data infrastructure for developing predictive analytics tools that will properly deliver on those expectations.

The goal of predictive analytics is to transform data into actionable insights. As it has matured as a discipline, we’ve finally seen the tangible benefits of big data emerge. The foremost examples can be found in healthcare. Analytical modeling of the vast amounts of data collected through electronic medical records systems is beginning to create positive solutions for healthcare system issues such as patient safety, physician and nurse burnout, and hospital staffing efficiency. Predictive analytics provides and continuously improves the foundation for biomarker-based patient stratification and therapeutic selection in oncology, predictive diagnostics for infectious diseases, critical care intervention, and many other clinical decision support systems.

One of the most meaningful use cases of predictive analytics in healthcare is the recent innovation of physiological avatars — digital twins of patients constructed from real-time accumulation of their physiological data. Driven by artificial intelligence (AI) and machine learning (ML), physiological avatars can, ideally, continuously evaluate a patient’s health in real-time and proactively alert them when their physiological state is trending in an unhealthy direction. With many advantages for patients, specifically, and healthcare, in general, they are, in many ways, the ultimate realization of the promise of “big data.”

A physiological avatar (PA) must accomplish several things to be more than a novelty. Specifically, it must:

• Provide a real-time, continuously updated digital twin of the patient
• Be able to provide a prognosis
• Evolve with the patient
• Be able to learn and improve as more data accumulates, both from the patient and external healthcare databases

The structure of a PA needs a continuously updated real-time source, which is where wearable digital technologies come into play. Smartwatches, sleep sensors, or skin-like healthcare patches can provide a continuous feed of several data types simultaneously, from heart rate and blood oxygenation percentage to blood sugar levels or the status of specific metabolites, such as lactate and ketones.

Periodic full physiological workups with one’s doctor can provide baselines and supplementary data, including more specific data inputs from imaging systems or specific diagnostic tests. When necessary, surgical implants can provide more acute monitoring, especially for patients with active disease states — poor cardiac, neurological, or oncology status, among others — or for those in remission from certain diseases. At the most basic level, periodic patient check-ins with online health system portals can help provide updates on mental status and descriptive symptomology.

All the physiological data contained in a PA isn’t very good unless it works to improve patient health and avert disease and other adverse health conditions. To that end, the data contained in a PA is compared and balanced against larger population-level physiological datasets that are accompanied by patient examination records and charts and optimally curated by healthcare professionals with valid experience in the clinical setting.

EMBS member Carlos Vega, from the Luxembourg Centre for Systems Biomedicine, has published several reports on the dangers of curating data for AI/ML-based predictive analytics without essential healthcare domain knowledge. The predictive models upon which these systems are based need to know which aspects of datasets are relevant versus those that may be confounding to the underlying analytical algorithms. This ensures AI/ML-based systems are fed the data they need to create prognostic and diagnostic power instead of being forced to provide outputs based on scant or incorrect data.

Predictive modeling is a statistically based methodology underlies physiological avatars’ diagnostic and prognostic power. As a patient’s physiological status is continuously built and adjusted in the form of their PA, it is constantly being evaluated by the AI, which drives the analytical systems of the PA. Based on an algorithmically based approach, each aspect of the physiological status of the PA is checked against the larger population-wide datasets using predictive modeling, which delivers a measure of how close to, or far from, a healthy state each physiological parameter is. Abnormal results would be further analyzed with subset algorithm-based predictive models that are more specific to individual physiological or disease states.

ML methods allow the predictive analytics and modeling within the PA technology to continuously learn and adapt to new data, such as increasing precision and volume in population-level datasets, as well as changes in the baseline physiological status of the patients themselves, which may occur due to aging or physiological changes brought on following injury or disease.

PAs should provide several incredible advantages to patients and healthcare providers (HCPs) when constructed and used correctly. First and foremost, among these is the increased precision a PA should be able to impart to clinical decision-support outputs and clinical decision-making. Without a PA to draw upon, HCPs work from single data points that are substantially isolated from the rest of a patient’s healthcare timeline.

The continuous data collection and curation inherent to PA technology will allow an HCP to analyze physiological trends leading up to changes in patient physiologies and compare those against the patient’s baseline status and larger population datasets. Most of this process will already be accomplished by the PA, leading up to the alert to the patient and HCP regarding new and potentially troubling changes in the patient’s physiological health.

Curation of ongoing patient data for real-time physiological benchmarking and personalization of the diagnostic decision space for ML models based on that data is itself, another set of advantages for PAs. It’s this overall richness of data quality that makes PAs potentially such a boon to the improvement of human health. This is further enhanced by integrating multiple tiers and types of data, including multi-omic inputs from genomics, proteomics, metabolomics, microbiomics, and others. Multi-omic inputs will allow patients to be continuously stratified within the context of population-level datasets, potentially leading to faster decision-making for the assignment of therapeutic and treatment plans.

Further, PAs can empower patients with their healthcare data, allowing them to monitor their health status actively and actively evaluate their therapeutic progress and response when necessary. The digital nature makes the process noninvasive and, for the most part, non-impactful for a patient’s daily routine. Further, because PAs are digital, they are not impacted by geographical barriers or the lack of healthcare systems. As such, they are potentially a robust tool for greatly improving access to quality healthcare in rural areas and underserved communities.

The ways in which PAs can potentially improve human health are staggering when considered in total. But, as an example, consider the potential for improving cardiac health. PAs could potentially impact cardiac health before or after adverse cardiac events (ACEs). Monitoring key cardiac indicators — pulse, oxygenation, body temperature, even blood pressure — is easily accomplished with basic wearables that many in the general population already use. This can be further supplemented with implanted monitors for those with declining cardiac health or following a major ACE to evaluate recovery.

Continuously captured data will begin to display a pattern, which PAs can subsequently monitor for changes that may indicate increasing probabilities for ACEs compared to the general population data. Following an ACE, the same systems can monitor patient recovery, and suggest whether treatments are working or a change in care is needed.

One can also see the potential power of PAs in the area of weight maintenance. PAs can help patients monitor weight variation and metabolic indicators, both of which can be employed in weight maintenance and diet plan evaluation. If feedback is available in real-time to help patients see positive results of their actions accumulating in their vital signs, this may also help improve patient adherence to exercise and diet regimes. Beyond these examples, the potential use cases are endless.

Unfortunately, there are still some barriers and disadvantages to fully realizing PAs within our healthcare system. Because of their heavy reliance on patient adherence to data tracking and curation, PAs can only be as useful as patients are willing to make them. By default, the advantage of making patients more active in their own healthcare decisions is a necessity to the successful implementation of PAs. As soon as the “advantage” stops being one, the system fails.

There is also, currently, a dearth of trained healthcare personnel to curate PA programs. Healthcare experts with domain knowledge are needed to inform the algorithms and ML frameworks that run the analytical processes within PAs, as well as to run the PA programs themselves.

Lastly, there is, currently, still a lack of robust ML-based models with malleable algorithms to adapt to an expanding and evolving diagnostic decision space. These are improving, however, and the complexity of biological knowledge is increasing as we continue to dive deeper into, and learn more about, human health and physiology. As diagnostic precision and depth constantly improve, the breadth and depth of data will continue to evolve and improve. As such, ML-based predictive modeling and analytics also need to constantly evolve. But the technologies that underlie the power of PAs — AI, ML, predictive analytics, predictive modeling, algorithmic engineering, etc. — are, themselves, still in their infancy.

It would be wonderful to state that we live in the age of physiological avatars. However, the fact remains, the technology is still in its nascent stages. That doesn’t mean there’s no benefit to be had just yet. On the contrary, there is, even now, great benefit to be had by getting involved in a PA program. And doing so will begin to add to the large public datasets, which are key to the success of these technologies. At the very least, it’s good to see that — due to the efforts of scientists, HCPs, engineers, and others involved in the creation of PAs — there is a technological epoch in this area to which we can look forward.

• Abeltino A, Bianchetti G, Serantoni C, Ardito CF, Malta D, De Spirito M, Maulucci G. Personalized metabolic avatar: a data driven model of metabolism for weight variation forecasting and diet plan evaluation. Nutrients. 14(17). (2022). DOI: 10.3390/nu14173520
• Sherry-Ann Brown. Principles for Developing Patient Avatars in Precision and Systems Medicine. Front. Genet. 6. (2016). DOI:10.3389/fgene.2015.00365
• Thomas A. Pearson, et al. Precision health analytics with predictive analytics and implementation research: JACC State-of-the-Art Review. 76(3). (2020). DOI:10.1016/j.jacc.2020.05.043
• Mijderwijk HJ, Steiger HJ. Predictive analytics in clinical practice: advantages and disadvantages. Acta Neurochir Suppl. 134. (2022). DOI: 10.1007/978-3-030-85292-4_30
• Carlos Vega. From Hume to Wuhan: An Epistemological Journey on the Problem of Induction in COVID-19 Machine Learning Models and Its Impact Upon Medical Research. IEEE Access. 9. 97243 (2021). DOI:10.1109/ACCESS.2021.3095222
• CArlos Vega, et al. Analysis: Flawed Datasets of Monkeypox Skin Images. Journal of Medical Systems. 47:37 (2023). DOI: 10.1007/s10916-023-01928-1
• Yang CC. Explainable artificial intelligence for predictive modeling in healthcare. J Healthc Inform Res. 6(2). (2022). DOI: 10.1007/s41666-022-00114-1
• Yeongjun Lee et al. Standalone real-time health monitoring patch based on a stretchable organic optoelectronic system. Sci. Adv. 7. eabg9180 (2021). DOI:10.1126/sciadv.abg9180
• Zhang Z. Predictive analytics in the era of big data: opportunities and challenges. Ann Transl Med. 8(4). (2020). DOI: 10.21037/atm.2019.10.97