Introduction

Pharmaceutical R&D increasingly depends on the ability to measure health not just in the clinic, but in the real world. While clinical biomarkers remain gold standards, behavioral and physiological signals captured through consumer-grade devices offer a scalable, complementary window into everyday health. At Evidation, we’ve spent more than a decade building wearable-based measures of health across a range of conditions - including infectious disease, immunology, and metabolic health. Most recently, we have developed a wearable-derived cardiometabolic score to quantify cardiometabolic health risk longitudinally and non-invasively, using passively collected data from consumer grade wearable devices.

Background and Motivation

Cardiometabolic disease refers to a broad constellation of interconnected conditions, many of which stem from long-term behaviors like inactivity, excess caloric intake, and poor sleep. Cardiometabolic dysfunction accumulates over time, but treatment often begins only after irreversible damage has occurred. Take type 2 diabetes, for example: a patient may be flagged as high risk and given an HbA1c test, which reveals an abnormally high average blood glucose, indicating a diagnosis of diabetes and associated beta cell death. Even common screening tools like BMI or waist circumference often serve as lagging indicators of an individual’s cardiometabolic health.

To address this, Evidation has developed a new method to measure the upstream physiological and behavioral patterns that precede cardiometabolic disease. Our approach combines artificial intelligence with wearable-derived data to capture early signals of cardiometabolic dysfunction, at scale and in real time.

This innovation presents several exciting new opportunities:

  • Screen and stratify individuals at elevated risk with greater sensitivity than traditional tools like BMI
  • Track treatment response to interventions, including GLP-1 therapies, via real-world behavioral and physiological shifts
  • Engage patients with behaviorally tailored content to support risk reduction and long-term health management

This score was developed using Evidation's connected, consented platform of hundreds of thousands of individuals who share longitudinal, real-world data - including wearables, self-reported outcomes (including validated PROs), EHR data, labs, and multimodal sensor inputs. 

Technical Approach

Our model utilized several key physiological, behavioral, and demographic variables as inputs, including:

  • Resting heart rate
  • Daily step count
  • Daily aerobic minutes
  • Sleep duration
  • Age
  • Gender

We used 90 day windows of a patient’s history across these variables in order to capture a representative longitudinal view of each individual's physical activity and behavior - much like the way an HbA1C captures the average level of glucose for an individual over the past 90 days. We then trained a machine learning model using these inputs to predict whether an individual’s BMI exceeded 30 (obese), using connected scale measurements. 

While BMI is an imperfect marker of metabolic health, it is strongly correlated with cardiometabolic dysfunction, and the hope was that the model would learn physiological, behavioral, and demographic characteristics that were correlated with high BMI, and therefore cardiometabolic health. Our development dataset consisted of ~1.2 million 90 day windows spanning 142,000 individuals. This dataset was then broken into a training set (used for model development) and a testing set (used for model validation).

Results

The trained model achieves an AUROC of 0.76 with an average precision of 0.72. These results are notable given that only consumer grade wearable devices were used to predict weight status. The model was also well calibrated, demonstrating an unbiased probability estimate that an individual was indeed obese.

The model calibration curve shows strong alignment between predicted probability and actual obesity prevalence - demonstrating that higher model output probabilities correspond closely to observed obesity rates across probability bins.

To further validate the model, we conducted additional validation exercises to explore the correlation between the newly derived score and an individual’s:

  • Risk of type 2 diabetes
  • Risk of hypertension
  • HbA1c

The  score exhibited significant discriminative abilities in differentiating those at high risk of type 2 diabetes and hypertension, and between individuals with varying HbA1c levels. Notably, individuals in the top quartile of cardiometabolic scores had nearly 9x higher incidence of diabetes than those in the bottom quartile.

Prevalence of Type 2 Diabetes (T2D) increases across cardiometabolic score quartiles, with individuals in the highest quartile (0.75–1.0) exhibiting a T2D prevalence nearly 9x greater than those in the lowest quartile (0.0–0.25).

Conclusion

The cardiometabolic score demonstrates how interactive real-world data (iRWD) can reveal subtle shifts in health that precede clinical diagnosis, offering a scalable, privacy-centric way to assess risk, monitor change, and engage individuals. This approach unlocks new opportunities to:

  • Pre-screen and stratify at risk individuals 
  • Track response to interventions, including GLP-1 therapies
  • Identify digital endpoints for decentralized trials
  • Support real-world patient engagement and activation

As digital biomarkers mature, tools like this offer new avenues for capturing treatment-relevant signals outside of traditional settings. Evidation is enabling life sciences partners to move beyond snapshots of health - capturing continuous, high-resolution insight into how everyday behavior shapes outcomes.

To learn more about  Evidation's cardiometabolic score, please contact us through: https://evidation.com/for-customers/contact-us

Have questions?

CONTACT US

Introduction

Pharmaceutical R&D increasingly depends on the ability to measure health not just in the clinic, but in the real world. While clinical biomarkers remain gold standards, behavioral and physiological signals captured through consumer-grade devices offer a scalable, complementary window into everyday health. At Evidation, we’ve spent more than a decade building wearable-based measures of health across a range of conditions - including infectious disease, immunology, and metabolic health. Most recently, we have developed a wearable-derived cardiometabolic score to quantify cardiometabolic health risk longitudinally and non-invasively, using passively collected data from consumer grade wearable devices.

Background and Motivation

Cardiometabolic disease refers to a broad constellation of interconnected conditions, many of which stem from long-term behaviors like inactivity, excess caloric intake, and poor sleep. Cardiometabolic dysfunction accumulates over time, but treatment often begins only after irreversible damage has occurred. Take type 2 diabetes, for example: a patient may be flagged as high risk and given an HbA1c test, which reveals an abnormally high average blood glucose, indicating a diagnosis of diabetes and associated beta cell death. Even common screening tools like BMI or waist circumference often serve as lagging indicators of an individual’s cardiometabolic health.

To address this, Evidation has developed a new method to measure the upstream physiological and behavioral patterns that precede cardiometabolic disease. Our approach combines artificial intelligence with wearable-derived data to capture early signals of cardiometabolic dysfunction, at scale and in real time.

This innovation presents several exciting new opportunities:

  • Screen and stratify individuals at elevated risk with greater sensitivity than traditional tools like BMI
  • Track treatment response to interventions, including GLP-1 therapies, via real-world behavioral and physiological shifts
  • Engage patients with behaviorally tailored content to support risk reduction and long-term health management

This score was developed using Evidation's connected, consented platform of hundreds of thousands of individuals who share longitudinal, real-world data - including wearables, self-reported outcomes (including validated PROs), EHR data, labs, and multimodal sensor inputs. 

Technical Approach

Our model utilized several key physiological, behavioral, and demographic variables as inputs, including:

  • Resting heart rate
  • Daily step count
  • Daily aerobic minutes
  • Sleep duration
  • Age
  • Gender

We used 90 day windows of a patient’s history across these variables in order to capture a representative longitudinal view of each individual's physical activity and behavior - much like the way an HbA1C captures the average level of glucose for an individual over the past 90 days. We then trained a machine learning model using these inputs to predict whether an individual’s BMI exceeded 30 (obese), using connected scale measurements. 

While BMI is an imperfect marker of metabolic health, it is strongly correlated with cardiometabolic dysfunction, and the hope was that the model would learn physiological, behavioral, and demographic characteristics that were correlated with high BMI, and therefore cardiometabolic health. Our development dataset consisted of ~1.2 million 90 day windows spanning 142,000 individuals. This dataset was then broken into a training set (used for model development) and a testing set (used for model validation).

Results

The trained model achieves an AUROC of 0.76 with an average precision of 0.72. These results are notable given that only consumer grade wearable devices were used to predict weight status. The model was also well calibrated, demonstrating an unbiased probability estimate that an individual was indeed obese.

The model calibration curve shows strong alignment between predicted probability and actual obesity prevalence - demonstrating that higher model output probabilities correspond closely to observed obesity rates across probability bins.

To further validate the model, we conducted additional validation exercises to explore the correlation between the newly derived score and an individual’s:

  • Risk of type 2 diabetes
  • Risk of hypertension
  • HbA1c

The  score exhibited significant discriminative abilities in differentiating those at high risk of type 2 diabetes and hypertension, and between individuals with varying HbA1c levels. Notably, individuals in the top quartile of cardiometabolic scores had nearly 9x higher incidence of diabetes than those in the bottom quartile.

Prevalence of Type 2 Diabetes (T2D) increases across cardiometabolic score quartiles, with individuals in the highest quartile (0.75–1.0) exhibiting a T2D prevalence nearly 9x greater than those in the lowest quartile (0.0–0.25).

Conclusion

The cardiometabolic score demonstrates how interactive real-world data (iRWD) can reveal subtle shifts in health that precede clinical diagnosis, offering a scalable, privacy-centric way to assess risk, monitor change, and engage individuals. This approach unlocks new opportunities to:

  • Pre-screen and stratify at risk individuals 
  • Track response to interventions, including GLP-1 therapies
  • Identify digital endpoints for decentralized trials
  • Support real-world patient engagement and activation

As digital biomarkers mature, tools like this offer new avenues for capturing treatment-relevant signals outside of traditional settings. Evidation is enabling life sciences partners to move beyond snapshots of health - capturing continuous, high-resolution insight into how everyday behavior shapes outcomes.

To learn more about  Evidation's cardiometabolic score, please contact us through: https://evidation.com/for-customers/contact-us

Have questions?

CONTACT US

Introduction

Pharmaceutical R&D increasingly depends on the ability to measure health not just in the clinic, but in the real world. While clinical biomarkers remain gold standards, behavioral and physiological signals captured through consumer-grade devices offer a scalable, complementary window into everyday health. At Evidation, we’ve spent more than a decade building wearable-based measures of health across a range of conditions - including infectious disease, immunology, and metabolic health. Most recently, we have developed a wearable-derived cardiometabolic score to quantify cardiometabolic health risk longitudinally and non-invasively, using passively collected data from consumer grade wearable devices.

Background and Motivation

Cardiometabolic disease refers to a broad constellation of interconnected conditions, many of which stem from long-term behaviors like inactivity, excess caloric intake, and poor sleep. Cardiometabolic dysfunction accumulates over time, but treatment often begins only after irreversible damage has occurred. Take type 2 diabetes, for example: a patient may be flagged as high risk and given an HbA1c test, which reveals an abnormally high average blood glucose, indicating a diagnosis of diabetes and associated beta cell death. Even common screening tools like BMI or waist circumference often serve as lagging indicators of an individual’s cardiometabolic health.

To address this, Evidation has developed a new method to measure the upstream physiological and behavioral patterns that precede cardiometabolic disease. Our approach combines artificial intelligence with wearable-derived data to capture early signals of cardiometabolic dysfunction, at scale and in real time.

This innovation presents several exciting new opportunities:

  • Screen and stratify individuals at elevated risk with greater sensitivity than traditional tools like BMI
  • Track treatment response to interventions, including GLP-1 therapies, via real-world behavioral and physiological shifts
  • Engage patients with behaviorally tailored content to support risk reduction and long-term health management

This score was developed using Evidation's connected, consented platform of hundreds of thousands of individuals who share longitudinal, real-world data - including wearables, self-reported outcomes (including validated PROs), EHR data, labs, and multimodal sensor inputs. 

Technical Approach

Our model utilized several key physiological, behavioral, and demographic variables as inputs, including:

  • Resting heart rate
  • Daily step count
  • Daily aerobic minutes
  • Sleep duration
  • Age
  • Gender

We used 90 day windows of a patient’s history across these variables in order to capture a representative longitudinal view of each individual's physical activity and behavior - much like the way an HbA1C captures the average level of glucose for an individual over the past 90 days. We then trained a machine learning model using these inputs to predict whether an individual’s BMI exceeded 30 (obese), using connected scale measurements. 

While BMI is an imperfect marker of metabolic health, it is strongly correlated with cardiometabolic dysfunction, and the hope was that the model would learn physiological, behavioral, and demographic characteristics that were correlated with high BMI, and therefore cardiometabolic health. Our development dataset consisted of ~1.2 million 90 day windows spanning 142,000 individuals. This dataset was then broken into a training set (used for model development) and a testing set (used for model validation).

Results

The trained model achieves an AUROC of 0.76 with an average precision of 0.72. These results are notable given that only consumer grade wearable devices were used to predict weight status. The model was also well calibrated, demonstrating an unbiased probability estimate that an individual was indeed obese.

The model calibration curve shows strong alignment between predicted probability and actual obesity prevalence - demonstrating that higher model output probabilities correspond closely to observed obesity rates across probability bins.

To further validate the model, we conducted additional validation exercises to explore the correlation between the newly derived score and an individual’s:

  • Risk of type 2 diabetes
  • Risk of hypertension
  • HbA1c

The  score exhibited significant discriminative abilities in differentiating those at high risk of type 2 diabetes and hypertension, and between individuals with varying HbA1c levels. Notably, individuals in the top quartile of cardiometabolic scores had nearly 9x higher incidence of diabetes than those in the bottom quartile.

Prevalence of Type 2 Diabetes (T2D) increases across cardiometabolic score quartiles, with individuals in the highest quartile (0.75–1.0) exhibiting a T2D prevalence nearly 9x greater than those in the lowest quartile (0.0–0.25).

Conclusion

The cardiometabolic score demonstrates how interactive real-world data (iRWD) can reveal subtle shifts in health that precede clinical diagnosis, offering a scalable, privacy-centric way to assess risk, monitor change, and engage individuals. This approach unlocks new opportunities to:

  • Pre-screen and stratify at risk individuals 
  • Track response to interventions, including GLP-1 therapies
  • Identify digital endpoints for decentralized trials
  • Support real-world patient engagement and activation

As digital biomarkers mature, tools like this offer new avenues for capturing treatment-relevant signals outside of traditional settings. Evidation is enabling life sciences partners to move beyond snapshots of health - capturing continuous, high-resolution insight into how everyday behavior shapes outcomes.

To learn more about  Evidation's cardiometabolic score, please contact us through: https://evidation.com/for-customers/contact-us

Have questions?

CONTACT US

Introduction

Pharmaceutical R&D increasingly depends on the ability to measure health not just in the clinic, but in the real world. While clinical biomarkers remain gold standards, behavioral and physiological signals captured through consumer-grade devices offer a scalable, complementary window into everyday health. At Evidation, we’ve spent more than a decade building wearable-based measures of health across a range of conditions - including infectious disease, immunology, and metabolic health. Most recently, we have developed a wearable-derived cardiometabolic score to quantify cardiometabolic health risk longitudinally and non-invasively, using passively collected data from consumer grade wearable devices.

Background and Motivation

Cardiometabolic disease refers to a broad constellation of interconnected conditions, many of which stem from long-term behaviors like inactivity, excess caloric intake, and poor sleep. Cardiometabolic dysfunction accumulates over time, but treatment often begins only after irreversible damage has occurred. Take type 2 diabetes, for example: a patient may be flagged as high risk and given an HbA1c test, which reveals an abnormally high average blood glucose, indicating a diagnosis of diabetes and associated beta cell death. Even common screening tools like BMI or waist circumference often serve as lagging indicators of an individual’s cardiometabolic health.

To address this, Evidation has developed a new method to measure the upstream physiological and behavioral patterns that precede cardiometabolic disease. Our approach combines artificial intelligence with wearable-derived data to capture early signals of cardiometabolic dysfunction, at scale and in real time.

This innovation presents several exciting new opportunities:

  • Screen and stratify individuals at elevated risk with greater sensitivity than traditional tools like BMI
  • Track treatment response to interventions, including GLP-1 therapies, via real-world behavioral and physiological shifts
  • Engage patients with behaviorally tailored content to support risk reduction and long-term health management

This score was developed using Evidation's connected, consented platform of hundreds of thousands of individuals who share longitudinal, real-world data - including wearables, self-reported outcomes (including validated PROs), EHR data, labs, and multimodal sensor inputs. 

Technical Approach

Our model utilized several key physiological, behavioral, and demographic variables as inputs, including:

  • Resting heart rate
  • Daily step count
  • Daily aerobic minutes
  • Sleep duration
  • Age
  • Gender

We used 90 day windows of a patient’s history across these variables in order to capture a representative longitudinal view of each individual's physical activity and behavior - much like the way an HbA1C captures the average level of glucose for an individual over the past 90 days. We then trained a machine learning model using these inputs to predict whether an individual’s BMI exceeded 30 (obese), using connected scale measurements. 

While BMI is an imperfect marker of metabolic health, it is strongly correlated with cardiometabolic dysfunction, and the hope was that the model would learn physiological, behavioral, and demographic characteristics that were correlated with high BMI, and therefore cardiometabolic health. Our development dataset consisted of ~1.2 million 90 day windows spanning 142,000 individuals. This dataset was then broken into a training set (used for model development) and a testing set (used for model validation).

Results

The trained model achieves an AUROC of 0.76 with an average precision of 0.72. These results are notable given that only consumer grade wearable devices were used to predict weight status. The model was also well calibrated, demonstrating an unbiased probability estimate that an individual was indeed obese.

The model calibration curve shows strong alignment between predicted probability and actual obesity prevalence - demonstrating that higher model output probabilities correspond closely to observed obesity rates across probability bins.

To further validate the model, we conducted additional validation exercises to explore the correlation between the newly derived score and an individual’s:

  • Risk of type 2 diabetes
  • Risk of hypertension
  • HbA1c

The  score exhibited significant discriminative abilities in differentiating those at high risk of type 2 diabetes and hypertension, and between individuals with varying HbA1c levels. Notably, individuals in the top quartile of cardiometabolic scores had nearly 9x higher incidence of diabetes than those in the bottom quartile.

Prevalence of Type 2 Diabetes (T2D) increases across cardiometabolic score quartiles, with individuals in the highest quartile (0.75–1.0) exhibiting a T2D prevalence nearly 9x greater than those in the lowest quartile (0.0–0.25).

Conclusion

The cardiometabolic score demonstrates how interactive real-world data (iRWD) can reveal subtle shifts in health that precede clinical diagnosis, offering a scalable, privacy-centric way to assess risk, monitor change, and engage individuals. This approach unlocks new opportunities to:

  • Pre-screen and stratify at risk individuals 
  • Track response to interventions, including GLP-1 therapies
  • Identify digital endpoints for decentralized trials
  • Support real-world patient engagement and activation

As digital biomarkers mature, tools like this offer new avenues for capturing treatment-relevant signals outside of traditional settings. Evidation is enabling life sciences partners to move beyond snapshots of health - capturing continuous, high-resolution insight into how everyday behavior shapes outcomes.

To learn more about  Evidation's cardiometabolic score, please contact us through: https://evidation.com/for-customers/contact-us

Have questions?

CONTACT US

Introduction

Pharmaceutical R&D increasingly depends on the ability to measure health not just in the clinic, but in the real world. While clinical biomarkers remain gold standards, behavioral and physiological signals captured through consumer-grade devices offer a scalable, complementary window into everyday health. At Evidation, we’ve spent more than a decade building wearable-based measures of health across a range of conditions - including infectious disease, immunology, and metabolic health. Most recently, we have developed a wearable-derived cardiometabolic score to quantify cardiometabolic health risk longitudinally and non-invasively, using passively collected data from consumer grade wearable devices.

Background and Motivation

Cardiometabolic disease refers to a broad constellation of interconnected conditions, many of which stem from long-term behaviors like inactivity, excess caloric intake, and poor sleep. Cardiometabolic dysfunction accumulates over time, but treatment often begins only after irreversible damage has occurred. Take type 2 diabetes, for example: a patient may be flagged as high risk and given an HbA1c test, which reveals an abnormally high average blood glucose, indicating a diagnosis of diabetes and associated beta cell death. Even common screening tools like BMI or waist circumference often serve as lagging indicators of an individual’s cardiometabolic health.

To address this, Evidation has developed a new method to measure the upstream physiological and behavioral patterns that precede cardiometabolic disease. Our approach combines artificial intelligence with wearable-derived data to capture early signals of cardiometabolic dysfunction, at scale and in real time.

This innovation presents several exciting new opportunities:

  • Screen and stratify individuals at elevated risk with greater sensitivity than traditional tools like BMI
  • Track treatment response to interventions, including GLP-1 therapies, via real-world behavioral and physiological shifts
  • Engage patients with behaviorally tailored content to support risk reduction and long-term health management

This score was developed using Evidation's connected, consented platform of hundreds of thousands of individuals who share longitudinal, real-world data - including wearables, self-reported outcomes (including validated PROs), EHR data, labs, and multimodal sensor inputs. 

Technical Approach

Our model utilized several key physiological, behavioral, and demographic variables as inputs, including:

  • Resting heart rate
  • Daily step count
  • Daily aerobic minutes
  • Sleep duration
  • Age
  • Gender

We used 90 day windows of a patient’s history across these variables in order to capture a representative longitudinal view of each individual's physical activity and behavior - much like the way an HbA1C captures the average level of glucose for an individual over the past 90 days. We then trained a machine learning model using these inputs to predict whether an individual’s BMI exceeded 30 (obese), using connected scale measurements. 

While BMI is an imperfect marker of metabolic health, it is strongly correlated with cardiometabolic dysfunction, and the hope was that the model would learn physiological, behavioral, and demographic characteristics that were correlated with high BMI, and therefore cardiometabolic health. Our development dataset consisted of ~1.2 million 90 day windows spanning 142,000 individuals. This dataset was then broken into a training set (used for model development) and a testing set (used for model validation).

Results

The trained model achieves an AUROC of 0.76 with an average precision of 0.72. These results are notable given that only consumer grade wearable devices were used to predict weight status. The model was also well calibrated, demonstrating an unbiased probability estimate that an individual was indeed obese.

The model calibration curve shows strong alignment between predicted probability and actual obesity prevalence - demonstrating that higher model output probabilities correspond closely to observed obesity rates across probability bins.

To further validate the model, we conducted additional validation exercises to explore the correlation between the newly derived score and an individual’s:

  • Risk of type 2 diabetes
  • Risk of hypertension
  • HbA1c

The  score exhibited significant discriminative abilities in differentiating those at high risk of type 2 diabetes and hypertension, and between individuals with varying HbA1c levels. Notably, individuals in the top quartile of cardiometabolic scores had nearly 9x higher incidence of diabetes than those in the bottom quartile.

Prevalence of Type 2 Diabetes (T2D) increases across cardiometabolic score quartiles, with individuals in the highest quartile (0.75–1.0) exhibiting a T2D prevalence nearly 9x greater than those in the lowest quartile (0.0–0.25).

Conclusion

The cardiometabolic score demonstrates how interactive real-world data (iRWD) can reveal subtle shifts in health that precede clinical diagnosis, offering a scalable, privacy-centric way to assess risk, monitor change, and engage individuals. This approach unlocks new opportunities to:

  • Pre-screen and stratify at risk individuals 
  • Track response to interventions, including GLP-1 therapies
  • Identify digital endpoints for decentralized trials
  • Support real-world patient engagement and activation

As digital biomarkers mature, tools like this offer new avenues for capturing treatment-relevant signals outside of traditional settings. Evidation is enabling life sciences partners to move beyond snapshots of health - capturing continuous, high-resolution insight into how everyday behavior shapes outcomes.

To learn more about  Evidation's cardiometabolic score, please contact us through: https://evidation.com/for-customers/contact-us

Have questions?

CONTACT US

Introduction

Pharmaceutical R&D increasingly depends on the ability to measure health not just in the clinic, but in the real world. While clinical biomarkers remain gold standards, behavioral and physiological signals captured through consumer-grade devices offer a scalable, complementary window into everyday health. At Evidation, we’ve spent more than a decade building wearable-based measures of health across a range of conditions - including infectious disease, immunology, and metabolic health. Most recently, we have developed a wearable-derived cardiometabolic score to quantify cardiometabolic health risk longitudinally and non-invasively, using passively collected data from consumer grade wearable devices.

Background and Motivation

Cardiometabolic disease refers to a broad constellation of interconnected conditions, many of which stem from long-term behaviors like inactivity, excess caloric intake, and poor sleep. Cardiometabolic dysfunction accumulates over time, but treatment often begins only after irreversible damage has occurred. Take type 2 diabetes, for example: a patient may be flagged as high risk and given an HbA1c test, which reveals an abnormally high average blood glucose, indicating a diagnosis of diabetes and associated beta cell death. Even common screening tools like BMI or waist circumference often serve as lagging indicators of an individual’s cardiometabolic health.

To address this, Evidation has developed a new method to measure the upstream physiological and behavioral patterns that precede cardiometabolic disease. Our approach combines artificial intelligence with wearable-derived data to capture early signals of cardiometabolic dysfunction, at scale and in real time.

This innovation presents several exciting new opportunities:

  • Screen and stratify individuals at elevated risk with greater sensitivity than traditional tools like BMI
  • Track treatment response to interventions, including GLP-1 therapies, via real-world behavioral and physiological shifts
  • Engage patients with behaviorally tailored content to support risk reduction and long-term health management

This score was developed using Evidation's connected, consented platform of hundreds of thousands of individuals who share longitudinal, real-world data - including wearables, self-reported outcomes (including validated PROs), EHR data, labs, and multimodal sensor inputs. 

Technical Approach

Our model utilized several key physiological, behavioral, and demographic variables as inputs, including:

  • Resting heart rate
  • Daily step count
  • Daily aerobic minutes
  • Sleep duration
  • Age
  • Gender

We used 90 day windows of a patient’s history across these variables in order to capture a representative longitudinal view of each individual's physical activity and behavior - much like the way an HbA1C captures the average level of glucose for an individual over the past 90 days. We then trained a machine learning model using these inputs to predict whether an individual’s BMI exceeded 30 (obese), using connected scale measurements. 

While BMI is an imperfect marker of metabolic health, it is strongly correlated with cardiometabolic dysfunction, and the hope was that the model would learn physiological, behavioral, and demographic characteristics that were correlated with high BMI, and therefore cardiometabolic health. Our development dataset consisted of ~1.2 million 90 day windows spanning 142,000 individuals. This dataset was then broken into a training set (used for model development) and a testing set (used for model validation).

Results

The trained model achieves an AUROC of 0.76 with an average precision of 0.72. These results are notable given that only consumer grade wearable devices were used to predict weight status. The model was also well calibrated, demonstrating an unbiased probability estimate that an individual was indeed obese.

The model calibration curve shows strong alignment between predicted probability and actual obesity prevalence - demonstrating that higher model output probabilities correspond closely to observed obesity rates across probability bins.

To further validate the model, we conducted additional validation exercises to explore the correlation between the newly derived score and an individual’s:

  • Risk of type 2 diabetes
  • Risk of hypertension
  • HbA1c

The  score exhibited significant discriminative abilities in differentiating those at high risk of type 2 diabetes and hypertension, and between individuals with varying HbA1c levels. Notably, individuals in the top quartile of cardiometabolic scores had nearly 9x higher incidence of diabetes than those in the bottom quartile.

Prevalence of Type 2 Diabetes (T2D) increases across cardiometabolic score quartiles, with individuals in the highest quartile (0.75–1.0) exhibiting a T2D prevalence nearly 9x greater than those in the lowest quartile (0.0–0.25).

Conclusion

The cardiometabolic score demonstrates how interactive real-world data (iRWD) can reveal subtle shifts in health that precede clinical diagnosis, offering a scalable, privacy-centric way to assess risk, monitor change, and engage individuals. This approach unlocks new opportunities to:

  • Pre-screen and stratify at risk individuals 
  • Track response to interventions, including GLP-1 therapies
  • Identify digital endpoints for decentralized trials
  • Support real-world patient engagement and activation

As digital biomarkers mature, tools like this offer new avenues for capturing treatment-relevant signals outside of traditional settings. Evidation is enabling life sciences partners to move beyond snapshots of health - capturing continuous, high-resolution insight into how everyday behavior shapes outcomes.

To learn more about  Evidation's cardiometabolic score, please contact us through: https://evidation.com/for-customers/contact-us

Have questions?

CONTACT US
Eve: Evidation's brand mark which is a yellow glowing orb

Introduction

Pharmaceutical R&D increasingly depends on the ability to measure health not just in the clinic, but in the real world. While clinical biomarkers remain gold standards, behavioral and physiological signals captured through consumer-grade devices offer a scalable, complementary window into everyday health. At Evidation, we’ve spent more than a decade building wearable-based measures of health across a range of conditions - including infectious disease, immunology, and metabolic health. Most recently, we have developed a wearable-derived cardiometabolic score to quantify cardiometabolic health risk longitudinally and non-invasively, using passively collected data from consumer grade wearable devices.

Background and Motivation

Cardiometabolic disease refers to a broad constellation of interconnected conditions, many of which stem from long-term behaviors like inactivity, excess caloric intake, and poor sleep. Cardiometabolic dysfunction accumulates over time, but treatment often begins only after irreversible damage has occurred. Take type 2 diabetes, for example: a patient may be flagged as high risk and given an HbA1c test, which reveals an abnormally high average blood glucose, indicating a diagnosis of diabetes and associated beta cell death. Even common screening tools like BMI or waist circumference often serve as lagging indicators of an individual’s cardiometabolic health.

To address this, Evidation has developed a new method to measure the upstream physiological and behavioral patterns that precede cardiometabolic disease. Our approach combines artificial intelligence with wearable-derived data to capture early signals of cardiometabolic dysfunction, at scale and in real time.

This innovation presents several exciting new opportunities:

  • Screen and stratify individuals at elevated risk with greater sensitivity than traditional tools like BMI
  • Track treatment response to interventions, including GLP-1 therapies, via real-world behavioral and physiological shifts
  • Engage patients with behaviorally tailored content to support risk reduction and long-term health management

This score was developed using Evidation's connected, consented platform of hundreds of thousands of individuals who share longitudinal, real-world data - including wearables, self-reported outcomes (including validated PROs), EHR data, labs, and multimodal sensor inputs. 

Technical Approach

Our model utilized several key physiological, behavioral, and demographic variables as inputs, including:

  • Resting heart rate
  • Daily step count
  • Daily aerobic minutes
  • Sleep duration
  • Age
  • Gender

We used 90 day windows of a patient’s history across these variables in order to capture a representative longitudinal view of each individual's physical activity and behavior - much like the way an HbA1C captures the average level of glucose for an individual over the past 90 days. We then trained a machine learning model using these inputs to predict whether an individual’s BMI exceeded 30 (obese), using connected scale measurements. 

While BMI is an imperfect marker of metabolic health, it is strongly correlated with cardiometabolic dysfunction, and the hope was that the model would learn physiological, behavioral, and demographic characteristics that were correlated with high BMI, and therefore cardiometabolic health. Our development dataset consisted of ~1.2 million 90 day windows spanning 142,000 individuals. This dataset was then broken into a training set (used for model development) and a testing set (used for model validation).

Results

The trained model achieves an AUROC of 0.76 with an average precision of 0.72. These results are notable given that only consumer grade wearable devices were used to predict weight status. The model was also well calibrated, demonstrating an unbiased probability estimate that an individual was indeed obese.

The model calibration curve shows strong alignment between predicted probability and actual obesity prevalence - demonstrating that higher model output probabilities correspond closely to observed obesity rates across probability bins.

To further validate the model, we conducted additional validation exercises to explore the correlation between the newly derived score and an individual’s:

  • Risk of type 2 diabetes
  • Risk of hypertension
  • HbA1c

The  score exhibited significant discriminative abilities in differentiating those at high risk of type 2 diabetes and hypertension, and between individuals with varying HbA1c levels. Notably, individuals in the top quartile of cardiometabolic scores had nearly 9x higher incidence of diabetes than those in the bottom quartile.

Prevalence of Type 2 Diabetes (T2D) increases across cardiometabolic score quartiles, with individuals in the highest quartile (0.75–1.0) exhibiting a T2D prevalence nearly 9x greater than those in the lowest quartile (0.0–0.25).

Conclusion

The cardiometabolic score demonstrates how interactive real-world data (iRWD) can reveal subtle shifts in health that precede clinical diagnosis, offering a scalable, privacy-centric way to assess risk, monitor change, and engage individuals. This approach unlocks new opportunities to:

  • Pre-screen and stratify at risk individuals 
  • Track response to interventions, including GLP-1 therapies
  • Identify digital endpoints for decentralized trials
  • Support real-world patient engagement and activation

As digital biomarkers mature, tools like this offer new avenues for capturing treatment-relevant signals outside of traditional settings. Evidation is enabling life sciences partners to move beyond snapshots of health - capturing continuous, high-resolution insight into how everyday behavior shapes outcomes.

To learn more about  Evidation's cardiometabolic score, please contact us through: https://evidation.com/for-customers/contact-us

Download app