The actualization of predictive healthcare, part of what I’ve referred to as Medicine 3.0, relies upon the effective integration of data from diverse sources, a task that is becoming increasingly possible thanks to artificial intelligence (AI). Specialized AI systems, such as deep learning neural networks, can process massive volumes of data, including images, speech, test results, and more. At the frontier of AI and medicine, sophisticated prediction models are being constructed to improve our ability to extract meaningful insights from the enormous data sets. These advancements allow for more accurate forecasts of an individual’s health journey and will facilitate the creation of precise and proactive intervention strategies. In an example of this potential, investigators Glaser et al. applied the powerful tool of AI to improve predictive models for all-cause mortality (ACM) based primarily on body composition data from DXA scans (also abbreviated DEXA), helping to reveal new insights on how body composition relates to health and lifespan.
An Incomplete Picture of Body Composition and Mortality
Body Mass Index (BMI) is a crude measure of body composition derived solely from the mass and height of an individual. While BMI is useful for looking at trends in the population level when other measurements are not available, this metric is limited when applied to individuals because it does not account for variations in muscle mass, bone density, or fat distribution.
In contrast, Dual-energy X-ray Absorptiometry, or DXA, provides a much more detailed breakdown of body composition. DXA uses low-dose X-rays to measure how two different types of energy pass through your body, ultimately providing an estimate of how the total mass of the body is divided across three components: lean tissue (muscle and organs), fat mass, and bone mineral content. In this way, DXA goes far beyond the simplicity of BMI, offering accurate metrics or calculations of bone mineral density, body fat percentage, total lean mass (e.g., ALMI and FFMI), total fat mass, BMI, and body composition in specific regions of the body (e.g., visceral fat mass, subcutaneous fat mass, and appendicular lean mass.)
In the field of mortality research, the relationship between body composition and mortality has been well established, largely thanks to the Health, Aging, and Body Composition study (Health ABC), but the sheer volume of data related to body composition has been prohibitive with respect to creating comprehensive predictive models of ACM.
Health ABC, a 16-year prospective cohort study, first explored the predictive power of baseline data, evaluating several body composition metrics derived from imaging data for their relationship to ACM (finding, for instance, that baseline strength is a predictor of ACM). In other words, though mortality data were collected over many years, predictive models were based solely on data from a single time point, as has historically been a common practice in creating such models. However, this approach is limited because it only provides a cross-sectional slice of an individual’s health status.
More recent studies have shifted their focus to analyzing longitudinal data, revealing the individual predictive value of changes in muscle mass and weight loss over an extended period of time. Models using longitudinal data enable researchers to observe changes associated with time; an important consideration when studying the effects of aging, yet such longitudinal studies have typically incorporated data from only one or a small handful of metrics. These studies have, in effect, collected over a longer period of time at the expense of the breadth of information included in predictive models. For example, previous longitudinal studies have shown isolated metrics like grip strength and walking speed to be associated with ACM, but have not created predictive models based on a collection of metrics that might cover many of the more nuanced changes in body composition over time.
Additionally, existing models for body composition and mortality are typically based on heavily filtered data. The raw output from a DXA scan is an image – and as the saying goes, “a picture is worth a thousand words.” Yet in practice, this rich source of information is distilled into a handful of general metrics (e.g., total lean body mass or fat mass), leaving out potentially valuable information from the original scan, such as where fat and muscle are distributed throughout the body.
AI Facilitates Use of More Data
Artificial intelligence can process enormous amounts of information at rates far faster than human brains. (Continuous advancements in AI make it impossible to give up-to-date numbers, but AI processing speeds are on the order of billions to trillions of times faster than those of humans.) This power makes it possible to analyze longitudinal data of broad scope, as well as to utilize the wealth of information in raw images from total-body DXA (TBDXA) scans, as recent advancements in AI have paved the way for direct incorporation of medical imaging data into ACM models. Glaser and colleagues thus sought to apply this tool, specifically using the technology of recurrent neural networks, to construct a more advanced and accurate model for predicting 10-year mortality probability based on comprehensive body composition data.
About the Study
For their study, Glaser et al. used data from the Health ABC cohort (n=3075, 48.4% female, 41.6% Black and 58.4% non-Hispanic White). At the time of recruitment, participants included in the Health ABC cohort were aged 70 to 79 years, had no difficulty with mobility required for normal living, and had not undergone treatment for cancer in the last three years. The researcher’s approach included the utilization of medical images from TBDXA, rather than the extracted data noted above. They then incorporated demographic information, such as race, sex, and age, as well as blood markers, including measurements of blood glucose, fasting glucose, blood insulin, fasting insulin, HgA1c, and interleukin-6. Additionally, they accounted for general fitness indicators like walking speed over various distances (3/4/6 meter, 20 m, and 400 m) and grip strength and self-reported questionnaire responses, which encompassed individuals’ abilities to engage in activities of daily living. Participants were continuously monitored through annual examinations and biannual questionnaires.
The investigators first analyzed cross-sectional data only (i.e., the single point in time look), creating three different models using different data inputs: clinical data only, TBDXA imaging data only, and a combination of both. The models’ accuracy in predicting 10-year survival was evaluated by calculating the “area under the receiver operating characteristic” (AUROC) – a common metric for determining the accuracy of predictive models – in which accuracy is reflected by AUROC values on a scale in which 0.5 is no better than chance and 1.0 is perfect predictive accuracy. The model that integrated both clinical data and TBDXA imaging (combined AUROC=0.71) outperformed models using each method individually (image-only AUROC=0.63; clinical data AUROC=0.69). The improvement in AUROC achieved by combining TBDXA imaging with clinical data suggests that the former provides unique information that complements the latter, highlighting the value of using imaging in mortality prediction models.
Subsequently, three longitudinal models were built to test the hypothesis that using TBDXA scans collected sequentially throughout the study period would outperform models using the same information from a single visit. All longitudinal models surpassed their single-record counterparts (combined AUROC=0.79; image-only AUROC=0.73; clinical data AUROC=0.76), demonstrating the importance of longitudinal assessment of multivariable data in predicting ACM. Further, comparing these three longitudinal models again demonstrates that accuracy is highest when using the combination of both raw scan images and clinical data. Limiting the model to only derived values from TBDXA and discarding other image information lessens the accuracy of the ACM model.
With Health Information, More is More
Glaser et al.’s cross-sectional models show that higher dimensionality in source data increases the accuracy of the resulting ACM predictions. Their longitudinal models show that accuracy is further improved by examining how multivariable data changes over time. In other words, when it comes to making predictions of health trajectories, more data means more accuracy, and more accuracy may in turn lead to a deeper understanding of the underlying biology linking risk variables with the outcome of interest, as well as to more targeted approaches for risk mitigation.
To illustrate this, let’s consider a simple readout of total body fat percentage distilled from TBDXA data. As I’ve said many times on the podcast, this body composition metric is certainly more useful in predicting risk of metabolic disease than, say, total body weight. But even better predictions can be made if we have the additional information of how that body fat is distributed between subcutaneous and visceral spaces. Visceral fat mass is much more predictive of metabolic dysfunction than subcutaneous fat mass, which tells us that the body responds to these different fat depots in distinct ways and that loss of fat mass is more imperative for those with substantial visceral fat than for those with substantial subcutaneous fat. The more detailed model enhances both our knowledge of biological foundations and our approaches to intervention. Now consider how much more refined that picture can become with raw TBDXA scans, which allow for a more granular look into the precise distribution of fat, muscle, and bone throughout the body.
Add to this the information on how these body composition details change over time. Consider two individuals with the same age and with similar baseline body composition data. Cross-sectional models would say they are equally likely to suffer a debilitating injury from a fall. But what if we look at the same two individuals five years later and find that one has lost considerably more muscle mass and bone density than the other over that span of time? Clearly, their risk profiles have diverged, but we would only know this if we had longitudinal data.
By understanding how the full picture of body composition is evolving over time, we can tailor interventions and lifestyle recommendations specific to each individual’s unique trajectory. A personalized approach has the potential to yield more effective and precise outcomes to address problem areas most susceptible to functional decline in the individual, promoting long-term health and well-being.
The Bottom Line
Glaser and colleagues’ research marks a notable step forward in the integration of AI in healthcare. Their study elegantly demonstrates the superiority of longitudinal models over those reliant on cross-sectional data. Furthermore, it emphasizes the valuable insights gained from using raw medical images, as opposed to data extracted from the images. The enhancement of predictive models for all-cause mortality offers a more precise and comprehensive understanding of a patient’s health journey, empowering clinicians to adopt more proactive intervention strategies. I anticipate that further work in this space will continue to refine models for risks of mortality and individual diseases and health concerns, paving the way for an era of improved healthcare outcomes through the synergy of AI and medical expertise.
For a list of all previous weekly emails, click here.