A documented research resource

CGM
Multimodal Resource

Glucose records.
Everyday context.

Continuous glucose monitoring, daily video diaries, and postvisit interviews from three studies at the University of Southern California. Explore what is available, how records connect, and what they can support.

Explore the resource View analysis examples

Source snapshot: October 12, 2026Documentation v1.0.0

SIGNAL + CONTEXTSCHEMATIC

Continuous glucose monitoringRepeated sensor observations

Calendar-date linkage

TextContextual accounts

AudioSpeech recordingsAcoustic features

VideoVisual contextDetection feasibility

Illustration only. Shared dates do not establish synchronized events.

86Wave-specific identifiersRepresented in media

1,387Media files33.62 hours of recordings

400,947Raw CGM observationsAcross three study waves

These counts describe different data layers. Wave-specific identifiers are not established as distinct people across studies; observed availability is not recruitment completion or recording adherence.

01 / DATASET

Three study waves.
Complementary perspectives.

The resource connects sensor records with participants’ accounts of CGM use, food, activities, and everyday experiences. Study designs and represented cohorts differ.

Source-study context and observed records by wave, October 12, 2026 snapshot.
Resource characteristic	G1	G2	G3
Recruitment context	Adult community health workers	Adult community health workers	Community-dwelling adults aged ≥35
Documented CGM schedule	20 unmasked days	10 masked + 10 unmasked days	10 masked + 10 unmasked days
Clinical-table records	20	30	24
Raw CGM observations	105,361	157,363	138,223
Daily media files	533	402	374
Postvisit media files	18	33	27
Observed daily keys	335	223	208

Daily keys use calendar dates in G1 and filename study-day labels in G2/G3. Protocol schedules describe intended collection, not verified wear time. Clinical tables, media, and registry enrollment have separate denominators.

Measure glucose

Repeated CGM observations provide a sensor record. All glucose values are expressed in mg/dL.

Record daily context

Smartphone diaries capture accounts prompted by CGM feedback and study questions.

Reflect after the visit

Semistructured interviews document experiences with monitoring and everyday activities.

02 / ANALYSIS EXAMPLES

What changes when
we define an analysis?

Descriptive demonstrations make selection and processing choices visible. These results characterize available records; they do not establish clinical prediction or cross-modal associations.

G2 + G3 DIARIES

Stable identifier counts.
Fewer eligible dates.

Daily CGM quality criteria retained 327 of 369 complete paired diary dates while retaining all 50 represented wave-specific identifiers.

369Paired dates

327CGM-eligible dates

Primary rule: at least 231 unique valid readings and a maximum gap of 60 minutes, including day boundaries. These are analyst-defined rules.

Open selection figure

G2 dates decrease from 223 observed media keys to 201 reference days, 199 complete paired days, and 177 CGM-eligible days. G3 decreases from 208 keys to 170 reference and paired days, then 150 eligible days. — Selection is measured at both the identifier and observation-date levels.

50IDs retained under the primary CGM rule

Allowing a maximum gap of 30, 60, or 120 minutes retains 320, 327, or 334 dates. Each rule keeps the same 50 IDs. A stable identifier denominator can conceal changing observation support.

Six panels show the distributions of within-identifier across-day IQRs for daily mean glucose, glucose SD, glucose CV, recording pitch median, pitch range, and mean HNR. The same eligible dates are used across panels, with 26 G2 and 22 G3 identifiers contributing at least two dates. — Each point summarizes one wave-specific identifier; these are descriptive distributions, not association estimates.

WITHIN-IDENTIFIER VARIATION

Shared dates.
Different measurements.

CGM and acoustic summaries can be described on the same eligible dates. Across-day IQRs require at least two dates, leaving 48 identifiers in this display.

CGM: Daily mean, sample SD, and coefficient of variation.
Audio: eGeMAPSv02 pitch and harmonics-to-noise summaries, weighted by detected phonation duration.
Text: LIWC-22 word categories describe the stored English text, with provenance strata kept separate.
Video: Sampled OpenFace detections describe processing feasibility.

Open variation figure

Interpretation boundaries

Language provenance. Original-versus-translated status remains undocumented for 361 of 371 candidate text days.

Time alignment. Shared calendar dates do not establish synchronized clocks or event-level correspondence.

Measurement validity. Finite acoustic features and accepted face detections do not establish validated psychological, clinical, or facial measurements.

03 / ACCESS & PROJECT

Have a research question
for this resource?

Requests are restricted to applicants affiliated with academic institutions. Contact Professor David Black to discuss the proposed use and applicable access conditions.

Contact Prof. David Black

davidbla@usc.edu

PROJECT INFORMATION

Research team & documentation

Huang-Cheng Chou · Bowen Yi · Nicholas Arce · Vu My · Ruth A Flores Cambara · Shrikanth Narayanan · David Black

Viterbi School of Engineering and Keck School of Medicine
University of Southern California

Manuscript: In preparation
Resource snapshot: October 12, 2026
Documentation release: v1.0.0 · October 13, 2026
Website source: View repository
Content provenance: Sources & interpretation notes

CGMMultimodal Resource

Three study waves.Complementary perspectives.