​A Digital Biomarker of Diabetes from Smartphone-Based Vascular Signals

Read the Full PapeR @ nature medicine  


Robert Avram, Jeffrey E. Olgin, Peter Kuhar, J. Weston Hughes, Gregory M Marcus, Mark J Pletcher, Kirstin Aschbacher*, Geoffrey H Tison*

Tison Lab | The Health eHeart Study


We developed a deep neural network algorithm to detect diabetes from vascular signals, called PPG or photoplethysmography, that can be measured with existing hardware in smart devices like smartphones and fitness trackers. This is the same technology that many wearable devices and hospitals use to measure heart rate.


Diabetes affects ~451 million people worldwide and nearly half are undiagnosed. Though its symptoms can go unnoticed for years before being diagnosed, diabetes impacts nearly every organ system causing substantial suffering and life-threatening diseases including heart attacks, stroke and kidney failure.



In a study published in Nature Medicine, we developed a deep neural network algorithm that can identify individuals with diabetes from signals captured using a standard smartphone camera. 


This study demonstrated that smartphone-measured PPG, analyzed with deep learning, can be used as a noninvasive digital biomarker of prevalent diabetes.





We developed a 39-layer convolutional deep neural network (DNN) to detect prevalent diabetes from smartphone PPG signals. 

The DNN takes a PPG waveform as the sole input and provides a score between 0 and 1, with higher scores suggesting greater likelihood of prevalent diabetes. The DNN accepts PPG waveforms ~21 seconds in duration, but shorter or longer durations can be padded or cropped, respectively. PPG waveforms were obtained by placing the index fingertip on the smartphone camera using the Azumio Instant Heart Rate smartphone application (Azumio, Inc). Changes in reflected light intensity recorded by the smartphone camera are interpreted as pulsatile blood volume change. PPG waveforms were colleced at either 100 or 120 Hz.


We assembled a large dataset of 2.6 million PPG recordings from 53,870 unique individuals to develop and validate the DNN.

Participants provided smartphone PPG measurements remotely during the study period and reported whether they had ever been diagnosed with diabetes by a healthcare provider. This "Primary Cohort" was randomly divided into training, development and test datasets in a ratio of ~7:1:2. Since many participants recorded >1 PPG recording, DNN performance was reported at the "recording-level" which treats each recording independently, and the "user-level" which averages the DNN score for all recordings provided by a user. User-level assessment was preferably reported when possible since clinical application calls for classifying a user as having diabetes or not.

We validated the DNN in three datasets: 1) the hold-out Test Dataset consisting of 11,313 people, 2) the "Contemporary Cohort" of 7,806 people and, 3) a real world "Clinic Cohort" of 181 people prospectively-enrolled from 3 clinics.

The Area Under the Receiver Operating Characteristic Curve (AUC), which is a general measure of algorithm performance, was 0.766 at the user-level in the Test Dataset validation. Validation results were similar in the other two validation datasets, demonstrating algorithm generalizability, showing that the algorithm performs similarly in populations and settings different from that in which it was trained.

DNN performance was similar in Clinic Cohort patients who had blood test-confirmed diabetes.

In a separate analysis, we examined the performance of the DNN in Clinic Cohort patients who had blood test-confirmed diabetes by either hemoglobin A1c or serum glucose. The DNN performance remained consistent at the recording-level. Furthermore, there was a positive correlation between the DNN score for diabetes and the blood test values of hemoglobin A1c and serum glucose.



We examined whether PPG was predictive of diabetes independently of other common predictors and co-occurring diseases. The PPG-based DNN score was strongly and independently predictive of diabetes.

We built nested logistic regression models in the Test Dataset with and without the inclusion of the DNN score. After adjustment for age, gender, race/ethnicity and body mass index, the DNN score remained independently predictive of diabetes. This was also true after adjusting for co-occurring comorbidities including high blood pressure, high cholesterol, coronary artery disease, prior heart attack, heart failure, peripheral vascular disease, prior stroke and sleep apnea. After addition of either basic demographics or comorbidities, the AUC increased to 0.830, suggesting improved predictive performance for diabetes.




We also examined the diagnostic odds-ratio for a positive DNN prediction across different Test Dataset strata. Performance was highest when users had >6 recordings and heart rate <100 beats/min.


We found that while heart rate and heart rate variability were important individual predictors of diabetes, the PPG data contains much of the information from heart rate and heart rate variability.

Heart rate and heart rate variability were individually significantly associated with diabetes. But both were attenuated in a multi-variable logistic regression model after adding the PPG DNN score, with heart rate variability becoming non-significant.

We also found that the DNN appears to have learned to ignore artifactual segments in the PPG waveform and that the majority of the PPG information seems to be encoded in the peak-to-peak PPG intervals.

We plotted activation maps for inner layers of the DNN to examine how the DNN deals with artifactual PPG segments. The activation patterns of some filters suggests that the DNN has learned to "ignore" regions of PPG artifact. 

A separate DNN trained using only peak-to-peak PPG intervals alone, without any PPG waveform data, achieved an AUC of 0.721, suggesting that the peak-to-peak intervals may contain the predominance of information relevant for PPG prediction of diabetes.



Our study demonstrates that smartphone-measured PPG, analyzed with deep learning, can serve as an independent and noninvasive digital biomarker of prevalent diabetes.

♦ Our validation of this digital biomarker in 3 cohorts demonstrated that the DNN performs consistently in external datasets, including in real-world clinical settings. 

♦ This work effectively expands the utility of the PPG modality since physicians do not currently interpret PPG in the context of diabetes.

♦ This tool could have particular impact in underserved populations and those out of reach from traditional medical care.


Identifying people at-risk for diabetes without clinic visits would significantly lower barriers to access given the ubiquity of smartphones, enabling measurement amongst many of the ~224 million people with undiagnosed diabetes.

The noninvasiveness and ease of measuring PPG with smart-devices make this tool widely scalable, while its painlessness makes it attractive for repeated testing.

Since the PPG biomarker is predictive independently of demographic and commonly co-occuring diseases, it can also supplement existing diabetes risk scores by capturing complementary vascular and autonomic information via PPG.


Read the Full PapeR @ nature medicine  

Email questions or inquiries here.