Truace
Subscribe
70/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Review of 16 studies finds hearing-loss prediction models at high risk of bias

Study title: Applying Machine Learning Models to Predict Occupational Noise-Induced Hearing Loss: A Systematic Review and Meta-analysis

Evidence type
Systematic review or meta-analysis
Sample
n = 29,094
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2713cite or check it

Study title: Applying Machine Learning Models to Predict Occupational Noise-Induced Hearing Loss: A Systematic Review and Meta-analysis

A systematic review and meta-analysis in Noise and Health pooled 16 studies, covering 29,094 workers, that used machine learning to predict occupational noise-induced hearing loss. Across 29 ML models the pooled AUC was 0.788, versus 0.811 for six logistic regression models.

Read the study abstract

OBJECTIVES: To map and critically appraise prediction models for occupational noise-induced hearing loss (ONIHL) using machine learning (ML) approaches and to evaluate reporting quality and risk of bias. METHODS: Seven databases were systematically searched. Reporting quality was assessed using the TRIPOD-AI checklist, and risk of bias and applicability were evaluated using the PROBAST tool. A meta-analysis was performed to evaluate ML model performance for predicting ONIHL using the area under the curve (AUC),…

Applying Machine Learning Models to Predict Occupational Noise-Induced Hearing Loss: A Systematic Review and Meta-analysis
"Medical Records" by Georgia National Guard, CC BY 2.0.

The problem

A systematic review of 16 studies on machine learning prediction of occupational noise-induced hearing loss found 56% at high risk of bias, none externally validated, and a pooled AUC of 0.788 for ML models versus 0.811 for logistic regression.

In brief

Nine of the 16 studies (56%) were rated at high risk of bias with the PROBAST tool, and 31% had high applicability concerns. All 16 relied on internal validation only, so the reported accuracy may not carry over to other workplaces. The authors conclude the evidence should be read with caution.

XGBoost and artificial neural networks scored highest in subgroup analysis (both 0.846), but with very high heterogeneity (I2 of 94.4% and 95.8%). The definition of hearing loss was a major source of the differences between studies.

Main points

  1. Seven databases were searched; 16 studies from 2015 to 2025 were included, covering 29,094 participants, mostly middle-aged men in Asian industries such as manufacturing and mining.
  2. Support vector machines were used in nine studies, random forests in seven, artificial neural networks in six, and decision trees and XGBoost in five each; only one study used a convolutional neural network.
  3. Age, noise level and exposure duration were the most common inputs; noise features such as kurtosis were rarely used.
  4. PROBAST rated 56% of studies high risk, 38% low and 6% unclear; the analysis domain was the main problem, with 44% rated high risk.
  5. Pooled AUC was 0.788 for 29 ML models and 0.811 for six logistic regression models, both with high heterogeneity (I2 of 93.4% and 87.4%).
  6. Studies defined hearing loss in at least seven different ways; heterogeneity was 0% for one definition and 95.8% for another.
  7. Average TRIPOD-AI reporting compliance across the 16 studies was 64%.

What the review found when it combined the studies

These are the review’s own pooled results, copied from its abstract. Truvace did not recalculate them.

  • Discrimination of machine learning models predicting occupational noise-induced hearing loss

    AUC 0.788 (95% CI 0.761 to 0.813)

    16 studies · 29,094 people

    The review’s own words
    The pooled AUC was 0.788 (95% CI: 0.761–0.813).

  • Discrimination of k-nearest neighbours models predicting occupational noise-induced hearing loss

    AUC 0.837 (95% CI 0.785 to 0.878)

    16 studies · 29,094 people

    The review’s own words
    KNN followed (AUC = 0.837; 95% CI: 0.785–0.878) with negligible heterogeneity ( I 2 = 3.6%).

  • Discrimination of logistic regression models predicting occupational noise-induced hearing loss

    AUC 0.811 (95% CI 0.743 to 0.865)

    16 studies · 29,094 people

    The review’s own words
    The pooled AUC was 0.811 (95% CI: 0.743–0.865).

Code checks every number above against the review’s text before it is shown. The short labels are an AI model’s reading of that text, so the review’s own sentence sits under each one. A pooled figure is only as strong as the studies behind it. How Truvace uses pooled results

The full story

The review followed PRISMA. From 2147 records, 1542 were screened after duplicates were removed, and 16 observational studies were included. Reporting quality was checked against the 27-item TRIPOD-AI checklist, with average compliance of 64%, and bias was assessed with PROBAST.

Pooled discrimination was similar for ML and logistic regression, with AUCs of 0.788 and 0.811. Individual studies often reported ML beating traditional models, including the ISO-1999 standard, but the pooled estimates did not show that advantage. In subgroups, XGBoost and ANN reached 0.846, KNN 0.837, random forest 0.808, SVM 0.792, decision trees 0.781 and AdaBoost 0.745.

The authors attribute bias mainly to small samples, too few events per predictor variable and weak handling of overfitting. All studies used hold-out or k-fold cross-validation; none validated on an independent cohort. They call for external validation, standard hearing-loss definitions and deep learning on noise time-domain features.

Sources

1 on file
  1. Noise and Healthdoi:10.4103/nah.nah_11_26
    Peer-reviewed2026-05-01

Talk it over

Four questions to start onWho carries this cost, and what would have to change to prevent it?

Start with a question

Review of 16 studies finds hearing-loss prediction models at high risk of biashttps://truvace.com/item/applying-machine-learning-models-to-predict-occupational-noise-induced-hearing-l

  1. Who carries this cost, and what would have to change to prevent it?

    The problem A systematic review of 16 studies on machine learning prediction of occupational noise-induced hearing loss found 56% at high risk of bias, none externally validated, and a pooled AUC of 0.788 for ML models versus 0.811 for logistic regression.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a systematic review or meta-analysis with a sample of 29,094, from Noise and Health.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat Heterogeneity was not fully explained by subgroup analyses, the number of studies per algorithm was small, and no included study was externally validated.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Health