Why this reading
More observations create more derivation decisions.
FDA will hold a workshop on 27 August 2026 about statistical considerations for digitally-derived endpoints. This review shows current use patterns and practical barriers.
For SP teams, continuous data require explicit quality flags, valid-wear rules, aggregation windows, software versions and traceability from raw signal to analysis value.
Reading order
Your 30-minute plan.
Predict why continuous data do not automatically produce better endpoints.
Read the Abstract, Introduction, Results and Discussion.
Identify statistical and data-standard questions.
Review fit-for-purpose implementation steps.
Design one sensor-endpoint derivation and QC pipeline.
Open-access sources
Current evidence and regulatory direction.
Brief background
From a sensor stream to one analysis value.
The review identified 48 studies across 36 countries. Thirty-eight studies used physiological sensor data, and 36 of these used continuous glucose monitoring.
Twelve studies reported sensor-based clinical outcomes such as physical activity, movement or sleep. Most devices were wrist-worn, and half combined sensor measures with other clinical assessments.
Continuous data may better reflect daily life, but device variation, software updates, insufficient training, non-wear and data-quality problems complicate analysis.
A programmer must convert raw timestamps into analysis-ready values using pre-specified quality checks, valid-day thresholds, aggregation windows, baseline rules and missing-data methods.
Key vocabulary
Fifteen terms worth retrieving.
| Term | 中文 | Meaning / use |
|---|---|---|
| digital health technology | 数字健康技术 | A sensor, software or connected system used to collect health data. |
| digitally-derived endpoint | 数字化衍生终点 | A trial endpoint calculated from data captured by a digital technology. |
| continuous glucose monitoring | 连续血糖监测 | Frequent measurement of interstitial glucose over time. |
| sensor-based functional outcome | 传感器功能性结局 | A passive measure of functioning, such as activity or sleep. |
| performance outcome | 表现性结局 | A measure produced while a participant performs a standardized task. |
| high-frequency data | 高频数据 | Measurements collected many times within a short period. |
| data aggregation window | 数据聚合窗口 | The time interval used to summarize raw observations. |
| wear-time compliance | 佩戴依从性 | How consistently participants use a device as required. |
| missingness mechanism | 缺失机制 | The process explaining why observations are absent. |
| fit-for-purpose | 适合预定用途 | Suitable and sufficiently validated for a defined context of use. |
| technical validation | 技术验证 | Evidence that a device measures its intended signal accurately. |
| analytical validation | 分析验证 | Evidence that algorithms correctly transform raw signals into measures. |
| clinical validation | 临床验证 | Evidence that a measure reflects a meaningful health concept. |
| software update | 软件更新 | A device or algorithm change that may alter generated data. |
| estimand alignment | 估计目标一致性 | Consistency between the endpoint, intercurrent events and treatment effect of interest. |
Useful phrases
Language for a digital-endpoint discussion.
- derive a clinically meaningful endpoint from continuous data - The team must derive a clinically meaningful endpoint from continuous data.
- pre-specify the aggregation algorithm - The analysis plan should pre-specify the aggregation algorithm.
- distinguish device non-wear from clinical inactivity - Programmers must distinguish device non-wear from clinical inactivity.
- define a minimum valid wear-time threshold - The protocol should define a minimum valid wear-time threshold.
- assess sensitivity to missing-data assumptions - The analysis should assess sensitivity to missing-data assumptions.
- maintain traceability from raw signal to endpoint - The pipeline must maintain traceability from raw signal to endpoint.
- lock the device and software version - The study should lock the device and software version where possible.
- combine sensor measures with complementary assessments - Sensor measures may be combined with complementary assessments.
- evaluate whether the measure is fit for purpose - The sponsor must evaluate whether the measure is fit for purpose.
- document every transformation step - The specification should document every transformation step.
Comprehension
Five questions.
- What types of sensor-derived data were most common?
- Why is CGM more mature than activity or sleep endpoints?
- What problems can software changes create?
- Why distinguish non-wear from true inactivity?
- Which derivation decisions should be pre-specified?
Retelling
Say it three times.
- 30 seconds · Review sample -> common technology -> main barrier.
- 45 seconds · Raw signal -> QC -> wear time -> aggregation -> endpoint.
- 60 seconds · Explain the impact on ADaM and TFL programming.
5-minute output task
Design a sensor-derived endpoint pipeline.
- Minute 1: Choose one endpoint.
- Minutes 2-3: Define raw input, quality flags, valid days, aggregation and baseline.
- Minute 4: Add version, timezone, missing-data, QC and traceability controls.
- Minute 5: State when the endpoint is analysis-ready.
One sentence to keep
Continuous data create richer evidence only when the endpoint algorithm, missing-data rules and software lineage are explicit, validated and reproducible.