Why this reading
A digital endpoint is not just another variable; it is a reproducible derivation pipeline.
Wearable and sensor data can arrive continuously, so the final analysis value depends on validity rules, time windows, aggregation, missingness handling, and device metadata.
FDA discussed these exact statistical and data-standard issues on 27 August 2026, while CDISC is preparing a September technology event centered on machine-readable, API-enabled standards workflows.
Reading order
Your 30-minute plan.
Identify physiological, functional and performance measurements.
Read CGM, activity, sleep, endpoint examples and implementation challenges.
Focus on statistical methods, data standards and CGM submission specifications.
Connect sensor pipelines to machine-readable, API-enabled standards automation.
Design one digital-endpoint derivation pipeline.
Open-access sources
Evidence, current regulatory discussion and standards-based technology direction.
Brief background
From raw signal to analysis endpoint.
The review included 48 recent clinical studies. Thirty-eight used sensor-based physiological data, with continuous glucose monitoring dominating that group.
Twelve studies used sensor-derived clinical outcomes such as physical activity and sleep. Common derived measures included time in glucose range, daily step counts, activity counts, total sleep time, and sleep efficiency.
These endpoints require explicit preprocessing and derivation choices. The statistical algorithm starts with the raw device stream, not with the final ADaM record.
FDA's August 27 workshop explicitly connected digitally-derived endpoints with statistical considerations, stronger data standards, innovative analytical methods, and continuous-glucose-monitoring technical specifications.
CDISC TechniCon frames the broader infrastructure shift: clinical research is moving from document-based workflows toward machine-readable and API-enabled data flows.
Key vocabulary
Fifteen terms for digital-endpoint programming.
| Term | 中文 | Meaning / use |
|---|---|---|
| digitally-derived endpoint | 数字化派生终点 | An endpoint computed from data captured by a digital health technology rather than a conventional site-based assessment. |
| digital health technology | 数字健康技术 | A sensor, software, connected device, or computing platform used to collect or process health-related data. |
| continuous glucose monitoring | 持续葡萄糖监测 | Repeated or continuous interstitial glucose measurement using a wearable sensor. |
| sensor-based functional outcome | 传感器功能结局 | A functional outcome derived passively from sensor data, such as activity or sleep. |
| performance outcome | 表现性结局 | A standardized task-based measure of what a participant can do. |
| high-frequency data | 高频数据 | Data recorded at short intervals, often producing many observations per participant per day. |
| time in range | 目标范围内时间 | The proportion of time a continuous measurement remains within a prespecified interval. |
| time above range | 高于目标范围时间 | The proportion of time a continuous measure exceeds a specified threshold. |
| time below range | 低于目标范围时间 | The proportion of time a continuous measure falls below a specified threshold. |
| data reduction | 数据降维 / 汇总 | Transforming dense raw sensor streams into interpretable summaries or endpoint variables. |
| wear-time rule | 佩戴时长规则 | A rule defining how much valid device use is required for an observation period to be analyzed. |
| missingness pattern | 缺失模式 | The structure and timing of unavailable measurements in a dataset. |
| fit-for-purpose | 适用于既定目的 | Suitable and sufficiently validated for the intended context of use. |
| machine-readable pipeline | 机器可读流程 | A workflow in which metadata, transformations, and outputs are represented in structured forms software can process. |
| endpoint derivation provenance | 终点派生谱系 | A traceable record connecting raw sensor data to processing steps, derived variables, and final endpoints. |
Useful phrases
Language for sensor-data and derivation discussions.
- turn a dense sensor stream into an analysis-ready endpoint - The pipeline must turn a dense sensor stream into an analysis-ready endpoint.
- define valid wear time before deriving the endpoint - Teams should define valid wear time before deriving the endpoint.
- separate raw acquisition from analysis derivation - The architecture separates raw acquisition from analysis derivation.
- preserve the lineage from device data to the final TFL - The workflow should preserve the lineage from device data to the final TFL.
- apply prespecified rules for missing sensor periods - Missing sensor periods should be handled using prespecified rules.
- summarize high-frequency measurements into clinically meaningful metrics - High-frequency measurements are summarized into clinically meaningful metrics.
- validate device-specific transformations independently - Device-specific transformations should be validated independently.
- treat the endpoint algorithm as part of the statistical specification - The endpoint algorithm should be treated as part of the statistical specification.
- design metadata for machine-readable reuse - Metadata should be designed for machine-readable reuse.
- make implementation choices visible to reviewers - Important implementation choices should be visible to reviewers.
Comprehension
Five questions.
- Why do continuous sensor measurements create a different programming problem from visit-based measurements?
- What does the widespread use of CGM suggest about digital-endpoint maturity?
- Why should wear-time and missing-data rules be specified before endpoint derivation?
- Why is FDA linking digitally-derived endpoints with both statistical methods and data standards?
- How could machine-readable metadata improve validation and reuse?
Retelling
Say it three times.
- 30 seconds · Define a digitally-derived endpoint and give two examples.
- 45 seconds · Raw sensor → validity rules → aggregation → endpoint → ADaM → TFL.
- 60 seconds · Explain why the endpoint algorithm belongs in the statistical specification.
5-minute output task
Design one digital-endpoint derivation pipeline.
- Minute 1: Choose glucose time-in-range, steps, activity, sleep, or AF burden.
- Minutes 2-3: Define frequency, wear-time, invalid-reading, aggregation, analysis-window and ADaM rules.
- Minute 4: Add device, algorithm, data-version, QC and missingness provenance.
- Minute 5: Explain how the endpoint can be reproduced independently.
One sentence to keep
When endpoints are derived from continuous sensors, statistical programming begins long before the final ADaM variable: the validity rules, time windows, aggregation algorithm, and provenance are part of the analysis itself.