Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 65 of the 65 references it cites were found in a registry or an index; none of the 28 sources its claims cite is checked for support yet.

Toward Generalizable Machine Learning Models in Speech, Language, and Hearing Sciences: Estimating Sample Size and Reducing Overfitting

Open access
Ghasemzadeh, Hamzeh, Hillman, Robert E., Mehta, Daryush D. · 2024 · Journal of Speech, Language, and Hearing Research

Abstract

Purpose: Many studies using machine learning (ML) in speech, language, and hearing sciences rely upon cross-validations with single data splitting. This study's first purpose is to provide quantitative evidence that would incentivize researchers to instead use the more robust data splitting method of nested k -fold cross-validation. The second purpose is to present methods and MATLAB code to perform power analysis for ML-based analysis during the design of a study. Method: First, the significant impact of different cross-validations on ML outcomes was demonstrated using real-world clinical data. Then, Monte Carlo simulations were used to quantify the interactions among the employed cross-validation method, the discriminative power of features, the dimensionality of the feature space, the dimensionality of the model, and the sample size. Four different cross-validation methods (single holdout, 10-fold, train–validation–test, and nested 10-fold) were compared based on the statistical power and confidence of the resulting ML models. Distributions of the null and alternative hypotheses were used to determine the minimum required sample size for obtaining a statistically significant outcome (5% significance) with 80% power. Statistical confidence of the model was defined as the probability of correct features being selected for inclusion in the final model. Results: ML models generated based on the single holdout method had very low statistical power and confidence, leading to overestimation of classification accuracy. Conversely, the nested 10-fold cross-validation method resulted in the highest statistical confidence and power while also providing an unbiased estimate of accuracy. The required sample size using the single holdout method could be 50% higher than what would be needed if nested k -fold cross-validation were used. Statistical confidence in the model based on nested k -fold cross-validation was as much as four times higher than the confidence obtained with the single holdout–based model. A computational model, MATLAB code, and lookup tables are provided to assist researchers with estimating the minimum sample size needed during study design. Conclusion: The adoption of nested k -fold cross-validation is critical for unbiased and robust ML studies in the speech, language, and hearing sciences. Supplemental Material: https://doi.org/10.23641/asha.25237045

Overview

CoChat Summary

## TL;DR

The paper argues that, in speech, language, and hearing ML studies with feature selection, nested 10-fold cross-validation is the most reliable of the four methods tested because it gives the highest statistical power and statistical confidence while remaining unbiased. It also provides Monte Carlo-based sample-size estimation code and lookup tables for planning ML studies.

## Key findings

  • In the real-world clinical example, single holdout and train-validation-test methods had unstable feature selection and biased or weak performance estimates, while nested 10-fold cross-validation produced the most consistent feature selection and an unbiased estimate of accuracy.
  • In simulated experiments, single holdout had “significant bias” and a heavy-tailed null distribution; nested k-fold cross-validation gave an unbiased estimate and a thinner upper tail for the null distribution.
  • The paper reports that, with 100 pairs, the probability that both selected features were correct was about 20% for single holdout but about 80% for nested 10-fold cross-validation.
  • The required sample size using single holdout “could be 50% higher” than with nested k-fold cross-validation, and statistical confidence for nested k-fold could be “as much as four times higher” than with single holdout.
  • For the nested 10-fold power analysis example, the paper reports that 89 pairs were needed for a statistically significant effect with \(d = 0.6\), \(m = 20\), and \(l = 2\); but 169 and 247 pairs were needed to reach 80% and 95% model confidence, respectively.

## Methods

  • Compared four cross-validation methods: single holdout, 10-fold, train-validation-test, and nested 10-fold cross-validation.
  • Used a real-world ambulatory voice-monitoring data set with 153 females with PVH and 136 female controls; the balanced analysis used 136 patients randomly selected, and analyses were repeated 500 times.
  • Ran Monte Carlo simulations for a balanced two-class problem, with Gaussian feature generation, wrapper-forward feature selection, and logistic regression; many simulated conditions were repeated 5,000 times, while the nested k-fold experiment was repeated 2,000 times.
  • Evaluated statistical power, null-hypothesis distribution characteristics, and statistical confidence; also derived and provided MATLAB code and lookup tables for required and recommended sample-size estimation.

## Limitations

  • The paper explicitly states that its sample-size equations were derived from simulated balanced data sets with underlying Gaussian distributions and equal discriminative power for selected features.
  • It also states that the main results were based on a binary classification problem, so the Experiment 3 sample-size results cannot be directly applied to other classification problems without separate simulation.
  • Evident from the text: the study only implemented nested \(k = 10\), so the generality of the findings to other values of \(k\) is not directly tested.

Model-written summary of a shortened version of the paper’s full text (it is longer than we read). Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1044/2023_jslhr-23-00273

Cites

65 of 65 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Visualization of data splitting for the four cross-validation methods studied: (A) single holdout, (B) k -fold ( k = 3), (C) train -validation -test, and (D) nested k -fold cross-validation ( k = 3).
Figure 1 Visualization of data splitting for the four cross-validation methods studied: (A) single holdout, (B) k -fold ( k = 3), (C) train -validation -test, and (D) nested k -fold cross-validation ( k = 3).From the PDF
Figure 2 (A) The effect of different cross-validations on the average and standard deviation of testing accuracy for different numbers of selected features. (B) The effect of sample size on testing accuracy estimated based on different cross-validations.
Figure 2 (A) The effect of different cross-validations on the average and standard deviation of testing accuracy for different numbers of selected features. (B) The effect of sample size on testing accuracy estimated based on different cross-validations.From the PDF
Figure 3 A comparison of statistical confidence of four cross-validation methods: (A) single holdout, (B) 10-fold, (C) train -validation -test, and (D) nested 10-fold. The horizontal axes represent the 48 dimensions in the feature space (the six voice measures listed under the panels with eight distributional characteristics per voice measure). Within each voice measure category on the plots, the distributional characteristics, in order, are the mean, median, standard deviation, interquartile range, fifth percentile, 95th percentile, skewness, and kurtosis. CPP = cepstral peak prominence; SPL = sound pressure level.
Figure 3 A comparison of statistical confidence of four cross-validation methods: (A) single holdout, (B) 10-fold, (C) train -validation -test, and (D) nested 10-fold. The horizontal axes represent the 48 dimensions in the feature space (the six voice measures listed under the panels with eight distributional characteristics per voice measure). Within each voice measure category on the plots, the distributional characteristics, in order, are the mean, median, standard deviation, interquartile range, fifth percentile, 95th percentile, skewness, and kurtosis. CPP = cepstral peak prominence; SPL = sound pressure level.From the PDF
Figure 4 Empirical estimation of the minimum required sample size as the value of n at the intersection between the upper bound of the H0 confidence interval (given α = .05) and the lower bound of the Ha confidence interval (given β = .2).
Figure 4 Empirical estimation of the minimum required sample size as the value of n at the intersection between the upper bound of the H0 confidence interval (given α = .05) and the lower bound of the Ha confidence interval (given β = .2).From the PDF
Figure 5 H0 characteristics: (A) mean and standard deviation of H0 accuracy and (B) upper bound of H0 CI (one-sided, α = .05). CI = confidence interval.
Figure 5 H0 characteristics: (A) mean and standard deviation of H0 accuracy and (B) upper bound of H0 CI (one-sided, α = .05). CI = confidence interval.From the PDF
Figure 6 Effect of the dimension of the feature space ( m ) on (A) mean H0 accuracy and (B) upper bound of H0 CI (one-sided, α = .05). CI = confidence interval.
Figure 6 Effect of the dimension of the feature space ( m ) on (A) mean H0 accuracy and (B) upper bound of H0 CI (one-sided, α = .05). CI = confidence interval.From the PDF
Figure 7 Comparison of the statistical power of the four cross-validation methods given varying simulated effect sizes and sample sizes. The portions of the figures that fall below the dashed black line correspond to underpowered cases: (A) single holdout, (B) 10-fold, (C) train -validation -test, and (D) nested 10-fold.
Figure 7 Comparison of the statistical power of the four cross-validation methods given varying simulated effect sizes and sample sizes. The portions of the figures that fall below the dashed black line correspond to underpowered cases: (A) single holdout, (B) 10-fold, (C) train -validation -test, and (D) nested 10-fold.From the PDF
Figure 8 Effect of different cross-validation methods on the statistical confidence of the trained model: (A) ( C 2,1 ) and (B) ( C 2,2 ).
Figure 8 Effect of different cross-validation methods on the statistical confidence of the trained model: (A) ( C 2,1 ) and (B) ( C 2,2 ).From the PDF
Figure 9 Effect of dimensionality of the feature space ( m ) on the statistical confidence of the trained model for nested 10-fold crossvalidation: (A) C 2 , 1 and (B) C 2 , 2 .
Figure 9 Effect of dimensionality of the feature space ( m ) on the statistical confidence of the trained model for nested 10-fold crossvalidation: (A) C 2 , 1 and (B) C 2 , 2 .From the PDF
Figure 10 Effect of the size of the feature space ( m ), the number of selected features ( l ), and effect size ( d ) on the required number of pairs ( nr ) for nested 10-fold cross-validation: (A) l = 2, (B) l = 3, and (C) l = 4.
Figure 10 Effect of the size of the feature space ( m ), the number of selected features ( l ), and effect size ( d ) on the required number of pairs ( nr ) for nested 10-fold cross-validation: (A) l = 2, (B) l = 3, and (C) l = 4.From the PDF
Figure 11 Effect of d and n on the confidence of the model generated using nested 10-fold cross-validation when m = 10.
Figure 11 Effect of d and n on the confidence of the model generated using nested 10-fold cross-validation when m = 10.From the PDF
Figure A1. The effect of an unbalanced data set (e.g., unequal number of patients and controls) on the estimation error of the minimum required sample size when Equations 4 -7 were used to estimate (A) the size of the smaller class and (B) an adjusted sample size, where the adjusted sample size was defined as the average of the sample sizes of the positive and negative classes. Please refer to the generalization of the results section for further details.
Figure A1. The effect of an unbalanced data set (e.g., unequal number of patients and controls) on the estimation error of the minimum required sample size when Equations 4 -7 were used to estimate (A) the size of the smaller class and (B) an adjusted sample size, where the adjusted sample size was defined as the average of the sample sizes of the positive and negative classes. Please refer to the generalization of the results section for further details.From the PDF
= 2 Figure A2. The effect of having two features with unequal discriminative power on the estimation error of the minimum required sample size when Equations 4 -7 were used with (A) the smaller effect size ( d 0 = d ) and (B) the average discriminatory power of the two features ( d 0 d γ d + 1 ( ) ). Please refer to the generalization of the results section for further details.
= 2 Figure A2. The effect of having two features with unequal discriminative power on the estimation error of the minimum required sample size when Equations 4 -7 were used with (A) the smaller effect size ( d 0 = d ) and (B) the average discriminatory power of the two features ( d 0 d γ d + 1 ( ) ). Please refer to the generalization of the results section for further details.From the PDF
Figure A3. Bull ' s eye nonlinear feature space: (A) with low level of noise (noise level = 2) and (B) with high level of noise (noise level = 4). Please refer to the generalization of the results section for further details.
Figure A3. Bull ' s eye nonlinear feature space: (A) with low level of noise (noise level = 2) and (B) with high level of noise (noise level = 4). Please refer to the generalization of the results section for further details.From the PDF

Verification

Claims the paper makes

0 Valid · 1 Not valid · 18 Not checked · 155 Own results and interpretations

122 more claims

Create a free account to see all 174 claims and to verify the 28 references still open.

See all claims — free account
  • It may state two separate claims as one. 35
  • The quote only partly states this claim. 5
  • Probably not a substantial claim. 12

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 63 references

Research works printed but not linked: 3

  1. Ghasemzadeh, H., & Searl, J. (2018, August 1–3). Modeling dynamics of connected speech in time and frequency domains with application to ALS [Paper presentation]. 11th International Conference on Voice Physiology and Biomechanics (ICVPB), East Lansing, MI, United States.
  2. Huang G. B. Mattar M. Berg T. & Learned-Miller E. (2008). Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Workshop on Faces in “Real-Life” Images: Detection Alignment and Recognition.
  3. Theodoridis S. & Koutroumbas K. (2009). Pattern recognition (4th ed.). Elsevier.

Cited by

See all 72

Metadata

Authors
Ghasemzadeh, H., Hillman, R. E., & Mehta, D. D.
Affiliations
Harvard University (US)
Venue
Journal of Speech, Language, and Hearing Research
Published
2024-03-11
Volume
67
Issue
3
Pages
753-781
ISSN
  • 1092-4388 (ISSN-L)
  • 1558-9102
OpenAlex
W4392040029
Cited by
72
References
63
Citations by year
2026: 29 · 2025: 36 · 2024: 7
FWCI
10.63
Type
journal article
Language
en
Open access
Licence
public-domain
Locations
  • Journal of Speech Language and Hearing Research (publishedVersion, public-domain)
  • PubMed Central (publishedVersion, public-domain)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by-nc-nd)
  • Digital Commons-TMC (Texas Medical Center) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • Massachusetts General Hospital
  • National Institutes of Health (DC015446)
  • National Institute on Deafness and Other Communication Disorders (K99DC021235, DC013017, DC015446, P50DC015446, T32DC013017)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
  • Health Sciences › Health Professions › Speech and Hearing › Voice and Speech Disorders
Keywords
sample size determination, statistical power, overfitting, machine learning, Monte Carlo simulation, research design
MeSH terms
Machine Learning, Hearing, Humans, Language, Speech, Sample Size

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Toward Generalizable Machine Learning Models in Speech, Language, and Hearing Sciences: Estimating Sample Size and Reducing Overfitting. https://app.cochat.ai/papers/10.1044/2023_jslhr-23-00273
The paper
Ghasemzadeh, H., Hillman, R. E., & Mehta, D. D. (2024). Toward Generalizable Machine Learning Models in Speech, Language, and Hearing Sciences: Estimating Sample Size and Reducing Overfitting. Journal of Speech, Language, and Hearing Research. https://doi.org/10.1044/2023_jslhr-23-00273

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.