Validation of the Diabetes Deep Neural Network Score for Diabetes Mellitus Screening

The Validation of the Diabetes Deep Neural Network Score (DNN score) for Screening for Type 2 Diabetes Mellitus (diabetes) is a single center, unblinded, observational study to clinically validating a previously developed remote digital biomarker, identified as the DNN score, to screen for diabetes. The previously developed DNN score provides a promising avenue to detect diabetes in these high-risk communities by leveraging photoplethysmography (PPG) technology on the commercial smartphone camera that is highly accessible. Our primary aim is to prospectively clinically validate the PPG DNN algorithm against the reference standards of glycated hemoglobin (HbA1c) for the presence of prevalent diabetes. Our vision is that this clinical trial may ultimately support an application to the Food and Drug Administration so that it can be incorporated into guideline-based screening.

Study Overview

• Status: Withdrawn
• Conditions: Diabetes
• Intervention / Treatment: Device: Application Validation

• Study Type: Interventional
• Phase: Not Applicable

Contacts and Locations

Study Locations

• United States
  • California
    • San Francisco, California, United States, 94143
      • University of California, San Francisco

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: 18 years and older (Adult, Older Adult)
• Accepts Healthy Volunteers: Yes

Description

Inclusion Criteria:

• Age > 18 years old
• Participants without a prior diagnosis of DM
• Participants with a recently measured HBA1c one month before enrollment or scheduled to undergo a HBA1c measurement within one month after enrollment
• Participants not scheduled for HBA1c and are willing to undergo a lab measured HBA1c
• Participants without risk factors for DM
• Participants with > 1 of the following risk factors for DM:
• Age > 40 years old
• Obesity (BMI > 30)
• Family history: Any first degree relative with a hx of DM
• Lifestyle risk factors (exercise, smoking, and sleep duration)
• Ownership of a smart phone
• Able to provide informed consent
• Willingness to provide PPG waveforms

Exclusion Criteria:

• Participants with a history of DM
• Participants with a prior HBA1c > 6.5%
• Inability to collect PPG signals (digit amputation, excessive tremors, etc)
• Lack of ownership of a smartphone
• Inability or unwillingness to consent and/or follow requirements of the study

Study Plan

How is the study designed?

Design Details

• Primary Purpose: Diagnostic
• Allocation: Non-Randomized
• Interventional Model: Parallel Assignment
• Masking: None (Open Label)

Number of Arms

2

Arms and Interventions

Participant Group / Arm	Intervention / Treatment
Experimental: Study Population; The investigators will conduct an electronic medical record (EMR) query of individuals in the University of California, San Francisco (UCSF) primary care clinics without a prior diagnosis of DM and who are undergoing, or who have recently undergone, a lab measured HBA1c before or after 1 month of enrollment. sample size estimation for testing the estimated AUROC in the validation sample vs. the null value of AUC 0.7. The investigators will target an enrollment of 5006 subjects in order to obtain a pre-specified AUROC 95% confidence interval width of 0.07 (i.e. AUROC = 0.76 [95%CI 0.725, 0.795]). The investigators assume that ~4% of the cohort will have undiagnosed diabetes based on national prevalence estimates.	Device: Application Validation; After creating accounts, participants in both groups will download the Azumio Instant Diabetes Test and provide a Photoplethysmography (PPG) waveforms by placing their index finger over their smartphone camera for 20 seconds to provide PPG waveform data for the study .
Experimental: Alternative Sample Group; The investigators also aim to perform a sensitivity analysis to estimate the DNN performance in a target general population without a diabetes diagnosis. The investigators will recruit patients from the UCSF EHR system without a history of diabetes, no prior HBA1c measured, and no history of known diabetic risk factors. The investigators will target an enrollment of 1000 subjects in order to obtain a pre-specified AUROC 95% confidence interval width of 0.18 (i.e. AUROC = 0.76 [95%CI 0.67, 0.85]). The investigators assume that ~3% of the cohort will have undiagnosed diabetes based on national prevalence estimates.	Device: Application Validation; After creating accounts, participants in both groups will download the Azumio Instant Diabetes Test and provide a Photoplethysmography (PPG) waveforms by placing their index finger over their smartphone camera for 20 seconds to provide PPG waveform data for the study .

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
The area under the receiver operating characteristic (AUROC) of the DNN Score as compared with one HBA1c measurement, based an average of two PPG measurements.	Participants will provide seven total PPG measurements by their own smartphone camera. After PPG measurements are obtained, the DNN algorithm will be deployed and be reported a as a DNN score. The investigators will assess the DNN performance by the the area under the receiver operating characteristic (AUROC) of the DNN Score as compared with the HBA1c based on the DNN score from an average of 2 PPG measurements.	PPG measurements and DNN score to be obtained within one month oh HBA1c measurement
The Sensitivity, Specificity, Positive Predictive Value, Negative Predictive Value of the DNN Score as compared with one HBA1c measurement based an average of two PPG measurements.	Participants will provide seven total PPG measurements by their own smartphone camera. After PPG measurements are obtained, the DNN algorithm will be deployed and be reported as a DNN score. The investigators will assess the DNN performance by the Sensitivity, Specificity, Positive Predictive Value, Negative Predictive Value of the DNN Score as compared with the HBA1c based on the DNN score from an average of 2 PPG measurements.	PPG measurements and DNN score to be obtained within one month oh HBA1c measurement
Assess the performance of the DNN score in different ethnicity and skin tones	The investigators will aim to recruit individuals of different races/ethnicities and skin tones to assess the performance of the DNN score in different races/ethnicities.	PPG measurements and DNN score to be obtained within one month oh HBA1c measurement

Secondary Outcome Measures

Outcome Measure	Measure Description	Time Frame
The area under the receiver operating characteristic (AUROC) of the DNN Score as compared with one HBA1c measurement based on > 2 PPG measurements.	Participants will provide seven total PPG measurements by their own smartphone camera. After PPG measurements are obtained, the DNN algorithm will be deployed and be reported a as a DNN score. The investigators will assess the DNN performance the area under the receiver operating characteristic (AUROC) of the DNN Score of > 2 PPG measurements as compared with the HBA1c.	PPG measurements and DNN score to be obtained within one month oh HBA1c measurement
The Sensitivity, Specificity, Positive Predictive Value, Negative Predictive Value of the DNN Score as compared with one HBA1c measurement based on >2 PPG measurements.	Participants will provide seven total PPG measurements by their own smartphone camera. After PPG measurements are obtained, the DNN algorithm will be deployed and be reported a as a DNN score. The investigators will assess the DNN performance by the Sensitivity, Specificity, Positive Predictive Value, Negative Predictive Value of the DNN Score of > 2 PPG measurements as compared with the HBA1c.	PPG measurements and DNN score to be obtained within one month oh HBA1c measurement
Retrain the DNN algorithm	By collecting PPG waveform data in patients with laboratory-confirmed diabetes, the investigators will be able to train the algorithm using the more specific diagnosis of laboratory-confirmed diabetes. The investigators will assess the performance of the DNN Score once retrained using HbA1c. The DNN will be trained using similar approaches as the investigators have previously published	Retraining to occur after complete collection of PPG measurements and HBA1c data. The investigators estimate this will occur one year after enrollment.

Collaborators and Investigators

• Sponsor: University of California, San Francisco
• Collaborators: Bristol-Myers Squibb; Azumio Inc.
• Investigators: Principal Investigator: Geoff Tison, MD, MPH, University of California, San Franscisco

Publications and helpful links

General Publications

• Avram R, Olgin JE, Kuhar P, Hughes JW, Marcus GM, Pletcher MJ, Aschbacher K, Tison GH. A digital biomarker of diabetes from smartphone-based vascular signals. Nat Med. 2020 Oct;26(10):1576-1582. doi: 10.1038/s41591-020-1010-5. Epub 2020 Aug 17.

Study record dates

Study Major Dates

• Study Start (Estimated): June 1, 2023
• Primary Completion (Actual): June 1, 2023
• Study Completion (Actual): April 1, 2025

Study Registration Dates

• First Submitted: March 10, 2022
• First Submitted That Met QC Criteria: March 21, 2022
• First Posted (Actual): March 31, 2022

Study Record Updates

• Last Update Posted (Actual): April 8, 2025
• Last Update Submitted That Met QC Criteria: April 4, 2025
• Last Verified: April 1, 2025

More Information

Terms related to this study

• Additional Relevant MeSH Terms: Endocrine System Diseases; Metabolic Diseases; Glucose Metabolism Disorders; Diabetes Mellitus
• Other Study ID Numbers: 21-35207

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: NO

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: Yes
• product manufactured in and exported from the U.S.: No