Can more frequent speech readings reduce sample size in CNS trials?
Often, yes. More readings per participant can reduce the noise in the estimate of each participant’s change, which can lower the number of participants a trial needs or shorten its follow‑up. That holds when the measure is reliable, responds to the change the trial targets, and has a variance structure the team understands.
For trial statisticians, clinical scientists and digital endpoint teams · Updated November 2026
Three conditions decide whether the gain is real
Frequent readings shrink one part of the variance a trial has to overcome. Whether that lowers the sample size depends on three things a sponsor can check in advance.
The reading is reliable
The same participant in the same state gives a similar reading from one session to the next. Without that, extra readings add noise faster than they remove it.
It responds to the change the trial targets
The reading moves with the clinical change the drug is meant to slow, on the scale the trial reports. Noise reduction on a measure that does not track the disease buys nothing.
Its variance is understood
The team knows how much of the spread is session‑to‑session noise, how much is true difference between participants, and how readings drift with practice or phone changes.
Why more readings can shrink the sample
The argument in plain words, with the papers that make it.
-
Two kinds of variance
Readings vary between participants, because people decline at different rates, and within each participant, because any single session is noisy. A trial comparing average change between arms has to see past both. More readings per participant reduce only the within‑participant part. For repeated‑measures trials analysed with summary means and ANCOVA, Frison and Pocock showed that each extra measurement adds less when measurements are highly correlated, and gave methods to compute the sample size.
-
Each participant’s change is a slope
A participant’s rate of change is estimated from their readings over time. The noise in that estimate falls as readings are added and as they spread across the follow‑up. Two readings a year apart give a rough slope; many readings across the same year give a tighter one.
-
Mixed models put the pieces together
A linear mixed model, or growth curve model, fits a slope for each participant and a mean slope per arm, and separates true differences in slopes from session noise. Its variance components feed the sample‑size calculation.
-
Where the gain stops
Once session noise is small next to the true spread in decline between participants, extra readings help little. Readings taken close together also share day‑to‑day fluctuation, so they count for less than independent ones. Measurement‑burst designs (Sliwinski, 2008) group several readings into short bursts spaced across the study: averaging within a burst cancels day‑to‑day fluctuation, and comparing bursts shows long‑term change apart from practice gains.
-
What frequent sampling has shown
In an Alzheimer’s prevention setting, Dodge and colleagues estimated that in‑home measures of walking speed and computer use, read against each person’s own baseline, needed far fewer participants to separate decline from stability than an annually assessed memory test. Brief smartphone cognitive tests averaged over many sessions reached reliability comparable to lab testing (Sliwinski et al.; Nicosia et al.).
-
Speech in particular
In ALS, an analysis presented at ISCTM in 2024 resampled cohorts of remote speech recordings and modelled each metric with growth curves; mean responsiveness held as cohorts got smaller, while its uncertainty grew (Kothare et al.). Stegmann and colleagues found low repeatability for many open‑source speech features, a reminder that reliability has to be shown for each measure.
Eight checks before a sample‑size gain goes in the protocol
Each one can erase the gain on paper if it is skipped.
Test‑retest reliability
Repeat sessions in a stable period, in the trial’s population and languages, give the within‑participant noise the calculation needs.
Sensitivity to change
The reading follows the clinical change the trial targets over the trial’s time frame, on the scale it reports.
Practice effects
Repeating a task can lift scores through familiarity alone. Goldberg et al. describe run‑in practice and alternate forms; Oravecz et al. model retest gains across bursts.
Adherence and missed sessions
The gain assumes the readings happen. Plan for realistic completion, and for missed sessions that cluster in the sickest participants.
Phone and software changes
Operating system updates, noise cancellation and new phones change the audio and can add drift that looks like disease change. The FDA’s 2023 DHT guidance asks sponsors to track such updates.
Timing of readings
Spread across the follow‑up, in bursts or on a fixed schedule, with time of day kept steady where it matters.
Pre‑specify in the SAP
The model, the reading schedule, how readings are summarised and how missing readings are handled are fixed before unblinding.
Simulate with pilot variance
Fit the planned model to pilot or prior data, then simulate the trial with realistic adherence, practice effects and drop‑out before committing to a smaller sample.
Speech readings between visits, beside the visit scale
A couple of minutes of speech can be recorded at home between clinic visits. That makes speech one practical way to add readings per participant.
More readings, same scale steps
A speech reading lands on the scale the trial already uses, at that scale’s own granularity. What it adds is frequency: readings between visits, beside the clinician’s rating.
Where the gain applies
The sample‑size gain applies to the endpoint whose readings are made more frequent: when a speech measure is the endpoint, or when frequent readings estimate the change the analysis tests, for example modelled jointly with the visit scale.
Variance from recordings you already hold
Speech or rating interviews recorded in a completed study can be read again to estimate session noise and between‑participant spread, the inputs a sample‑size simulation needs.
Home, site and existing recordings
Readings can come from short home sessions, from recordings sites already make at visits, and from audio a trial already holds, joined into one series per participant.
Readings between visits, and the variance to plan with
Cephalgo reads speech between clinic visits, from home sessions and site recordings, on or beside the scale the trial already uses. From a completed study, Cephalgo returns the within‑ and between‑participant variance components of its readings, the inputs your statistician needs for a sample‑size simulation for the next protocol. Readings stay stable across phone and software changes, and each one carries its analysis version. Re‑analysis pilot › · Device changes › · Validation approach › · Speech biomarker companies ›
Papers and guidance behind this page
Dodge HH, Zhu J, Mattek NC, Austin D, Kornfeld J, Kaye JA. Use of high‑frequency in‑home monitoring data may reduce sample sizes needed in clinical trials. PLoS One. 2015;10(9):e0138095. doi:10.1371/journal.pone.0138095
Compares sample sizes for Alzheimer’s prevention trials using an annual memory test and person‑specific thresholds from continuous in‑home monitoring.
Frison L, Pocock SJ. Repeated measures in clinical trials: analysis using mean summary statistics and its implications for design. Statistics in Medicine. 1992;11(13):1685‑1704. doi:10.1002/sim.4780111304
Sample‑size methods for repeated‑measures trials, including the value of more than one measurement per period.
Sliwinski MJ. Measurement‑burst designs for social health research. Social and Personality Psychology Compass. 2008;2(1):245‑261. doi:10.1111/j.1751-9004.2007.00043.x
How bursts of intensive assessment can improve detection of long‑term change within each person.
Sliwinski MJ, Mogle JA, Hyun J, Munoz E, Smyth JM, Lipton RB. Reliability and validity of ambulatory cognitive assessments. Assessment. 2018;25(1):14‑30. doi:10.1177/1073191116643164
Averaging brief smartphone tests across sessions gave reliability comparable to lab assessments.
Nicosia J, Aschenbrenner AJ, Balota DA, et al. Unsupervised high‑frequency smartphone‑based cognitive assessments are reliable, valid, and feasible in older adults at risk for Alzheimer’s disease. J Int Neuropsychol Soc. 2023;29(5):459‑471. doi:10.1017/S135561772200042X
Reliability, validity and adherence of repeated unsupervised testing at home in older adults.
Goldberg TE, Harvey PD, Wesnes KA, Snyder PJ, Schneider LS. Practice effects due to serial cognitive assessment: implications for preclinical Alzheimer’s disease randomized controlled trials. Alzheimers Dement (Amst). 2015;1(1):103‑111. doi:10.1016/j.dadm.2014.11.003
Practice gains in serial testing and three ways to reduce them.
Oravecz Z, Harrington KD, Hakun JG, et al. Accounting for retest effects in cognitive testing with the Bayesian double exponential model via intensive measurement burst designs. Front Aging Neurosci. 2022;14:897343. doi:10.3389/fnagi.2022.897343
A model that separates retest gains from change across bursts.
Stegmann GM, Hahn S, Liss J, et al. Repeatability of commonly used speech and language features for clinical applications. Digital Biomarkers. 2020;4(3):109‑122. doi:10.1159/000511671
Repeatability of open‑source speech features in healthy controls, ALS and suspected frontotemporal dementia.
Kothare H, Neumann M, Ramanarayanan V. Relationship between sample size and responsiveness of speech‑based digital biomarkers in ALS. ISCTM 20th Annual Scientific Meeting, 2024. Abstract (isctm.org)
Resampled cohorts of remote speech recordings in ALS, with growth curve models.
U.S. Food and Drug Administration. Digital Health Technologies for Remote Data Acquisition in Clinical Investigations. Final guidance, December 2023.
Fitness for purpose of digital measures, and how sponsors should record and handle technology and phone updates.
Speech readings and sample size
Can more frequent speech readings reduce sample size in a CNS trial?
They can, when the reading is reliable, responds to the change the trial targets and has a well‑understood variance structure. Extra readings reduce the noise in each participant’s estimated change, which can lower the number of participants needed or shorten follow‑up. The size of the gain should be estimated from pilot or prior data before it enters a sample‑size calculation.
Why do more readings per participant help?
Each participant’s change is estimated from their readings over time. Session noise averages out as readings are added and spread across the follow‑up, so each participant’s slope is estimated more tightly. True differences in decline between participants do not shrink with more readings, and they set a floor on the gain.
What limits the gain from frequent readings?
Practice effects, missed sessions, readings bunched close in time, shared day‑to‑day fluctuation, and phone or software changes. Each can add bias, or leave noise that extra readings do not remove.
How should a sponsor estimate sample size for a design with frequent readings?
Take variance components from a pilot or a completed study, fit the planned mixed model, and simulate the trial with realistic adherence, practice effects and drop‑out. Pre‑specify the model, the reading schedule and the handling of missing readings in the statistical analysis plan.
Do frequent speech readings replace the clinic visit scale?
No. Speech readings sit between visits, beside the scale the trial already uses, and land on that scale at its own granularity. What they add is frequency. The sample‑size gain applies to the endpoint built from the frequent readings, when a speech measure is the endpoint or when frequent readings estimate the change the analysis tests.
How does Cephalgo support sample-size planning?
Cephalgo reads speech between clinic visits and, from a completed study, returns the variance components of its readings so the sponsor’s statistician can compute a sample‑size estimate. Readings stay stable across phone and software changes.