Truace
Subscribe
70/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Review of 23 PTSD prediction studies finds none had low risk of bias

Study title: Machine learning algorithms for predicting PTSD: a systematic review and meta-analysis

Evidence type
Systematic review or meta-analysis
Sample
n = 23
Sources
1 sourcePeer-reviewed
Published
3 min read
Truvace record
TRV-2026-2753cite or check it

Study title: Machine learning algorithms for predicting PTSD: a systematic review and meta-analysis

A systematic review and meta-analysis of machine learning models that predict post-traumatic stress disorder pooled results from 23 studies published from 2018 to December 2023. Using the PROBAST bias tool, the authors rated 48% of the studies (11) at high risk of bias and the other 12 as unclear; none were rated low.

Read the study abstract

Abstract This study aimed to compare and evaluate the prediction accuracy and risk of bias (ROB) of post-traumatic stress disorder (PTSD) predictive models. We conducted a systematic review and random-effect meta-analysis summarizing predictive model development and validation studies using machine learning in diverse samples to predict PTSD. Model performances were pooled using the area under the curve (AUC) with a 95% confidence interval (CI). Heterogeneity in each meta-analysis was measured using I2. The risk…

Machine learning algorithms for predicting PTSD: a systematic review and meta-analysis
"Men Therapy Group - Scene from eMANcipation" by Philipp Müller-Dorn, CC BY-SA 4.0.

The problem

A systematic review of 23 machine learning studies predicting PTSD found none had low risk of bias (11 high, 12 unclear), and in medical-trauma populations external validation of tree-based models pooled an AUC of 0.591.

In brief

Tree-based models were the most commonly used algorithm and scored highest on pooled AUC: 0.96 for firefighters, 0.935 for alcohol-related stress, 0.861 for sexual or physical trauma, 0.808 for medical trauma, 0.771 for natural disasters and 0.745 for military incidents. Heterogeneity was very high, with I2 values often above 90%.

Only two studies tested their models on outside data. Where external validation could be pooled, performance fell: tree-based models for medical trauma reached 0.808 internally but 0.591 externally. The authors say this limits how far the results apply to real clinical settings.

Main points

  1. The review searched PubMed, Scopus and Web of Science for English-language studies from 2018 to 14 December 2023, screened 1,299 articles after removing duplicates, and included 18 studies plus 5 found through reference lists and Google, totalling 23.
  2. Fifteen of the 23 studies (65%) came from the United States and four from China; 14 used tree-based models, eight used linear models and five each used SVMs and neural networks.
  3. In war and military populations, tree-based models pooled an AUC of 0.745 against 0.615 for neural networks; in natural-disaster populations SVMs pooled only 0.593 against 0.771 for tree-based models.
  4. The analysis domain carried the most bias: 43% of studies (10) were rated high risk, 52% unclear and one low, with small samples and excluded participants named as main causes.
  5. Heterogeneity was often extreme, for example I2 of 99.98% for firefighters, though sensitivity analyses cut it to 0% for tree-based models in war populations.
  6. Most studies did not report participants' racial background, which the authors say limits understanding of how well the models fit different groups.
  7. The authors recommend following AI/ML reporting standards and PROBAST guidelines, and prioritising external validation before these models are used clinically.

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.

  • PTSD prediction by tree-based models in people with alcohol-related stress

    AUC 0.935 (95% CI 0.914 to 0.956)

    The review’s own words
    Similarly, in populations with alcohol-related stress, these models produced a pooled AUC of 0.935 (95% CIs 0.914–0.956, I 2 = 97.55%).

  • PTSD prediction by ensemble methods in the general population

    AUC 0.841 (95% CI 0.783 to 0.898)

    The review’s own words
    In general PTSD populations, the meta-analysis included SVM, ensemble methods, and linear models, recording pooled AUCs of 0.824 (95% CIs 0.778–0.871, I 2 = 0%), 0.841 (95% CIs 0.783–0.898, I 2 = 0%), and 0.768 (95% CIs 0.608–0.928, I 2 = 60.06%), respectively.

  • PTSD prediction by linear models in people with medical trauma

    AUC 0.828 (95% CI 0.717 to 0.939)

    The review’s own words
    In the context of medical trauma, analyses were conducted for two algorithms, revealing pooled AUCs of 0.828 (95% CIs 0.717–0.939, I 2 = 98.18%) for linear models and 0.808 (95% CIs 0.761–0.855, I 2 = 95.25%) for tree-based models.

  • PTSD prediction by tree-based models in people with war or military experience

    AUC 0.745 (95% CI 0.572 to 0.917)

    The review’s own words
    Specifically, tree-based models achieved a pooled AUC of 0.745 (95% CIs 0.572–0.917, I 2 = 97.31%), compared to 0.615 (95% CIs 0.552–0.768, I 2 = 0%) for neural networks.

  • PTSD prediction by tree-based models in people with sexual or physical trauma

    AUC 0.861 (95% CI 0.723 to 1)

    The review’s own words
    For individuals with experiences of sexual or physical trauma, the employment of tree-based models resulted in a pooled AUC of 0.861 (95% CIs 0.723–1, I 2 = 96.78%).

  • PTSD prediction by tree-based models in people with medical trauma

    AUC 0.808 (95% CI 0.761 to 0.855)

    The review’s own words
    In the context of medical trauma, analyses were conducted for two algorithms, revealing pooled AUCs of 0.828 (95% CIs 0.717–0.939, I 2 = 98.18%) for linear models and 0.808 (95% CIs 0.761–0.855, I 2 = 95.25%) for tree-based models.

  • PTSD prediction by tree-based models in firefighters

    AUC 0.96 (95% CI 0.918 to 1)

    The review’s own words
    Among firefighters, tree-based models achieved the highest pooled AUC across all examined demographics, at 0.96 (95% CIs 0.918–1, I 2 = 99.98%).

  • PTSD prediction by tree-based models after natural disasters

    AUC 0.771 (95% CI 0.755 to 0.787)

    The review’s own words
    Tree-based models led with a pooled AUC of 0.771 (95% CIs 0.755–0.787, I 2 = 90.41%), closely followed by linear models with an AUC of 0.768 (95% CIs 0.764–0.771, I 2 = 78.76%).

  • PTSD prediction by neural networks in people with war or military experience

    AUC 0.615 (95% CI 0.552 to 0.768)

    The review’s own words
    Specifically, tree-based models achieved a pooled AUC of 0.745 (95% CIs 0.572–0.917, I 2 = 97.31%), compared to 0.615 (95% CIs 0.552–0.768, I 2 = 0%) for neural networks.

  • External validation of tree-based models predicting PTSD in people with medical trauma

    AUC 0.591 (95% CI 0.47 to 0.712)

    The review’s own words
    External validation of tree-based models in this demographic reported a pooled AUC of 0.591 (95% CIs 0.47–0.712, I 2 = 93.61%).

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 authors pooled area-under-the-curve scores separately for each trauma population and for each algorithm family, using random-effects models, and measured heterogeneity with I2. A pool needed at least two studies, so single-study results such as 0.75 for a random forest in commercially insured adults were reported without pooling. Two reviewers screened and assessed bias independently.

Reported accuracy varied widely by group. Pooled AUCs for tree-based models ranged from 0.745 for military incidents to 0.96 for firefighters. In general PTSD populations, ensemble methods (0.841) and SVMs (0.824) outperformed linear models (0.768). Individual studies reported AUCs from 0.46 to 0.99 depending on model and validation set.

External validation was rare: only two studies did it, both with tree-based models. In medical trauma the external pooled AUC of 0.591 was far below the internal 0.808, and a military study's external AUC of 0.74 was also below its development results. The authors attribute this to overfitting, small samples and differences between datasets.

The authors also note that a model needing 50 or more predictors may be impractical, so clinicians might prefer traditional tools such as the PCL-5. They add that testing effects on patient outcomes would need interventional trials and impact assessments.

Sources

1 on file
  1. BMC Medical Informatics and Decision Makingdoi:10.1186/s12911-024-02754-2
    Peer-reviewed2025-01-21

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 23 PTSD prediction studies finds none had low risk of biashttps://truvace.com/item/machine-learning-algorithms-for-predicting-ptsd-a-systematic-review-and-meta-ana

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

    The problem A systematic review of 23 machine learning studies predicting PTSD found none had low risk of bias (11 high, 12 unclear), and in medical-trauma populations external validation of tree-based models pooled an AUC of 0.591.

    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 23, from BMC Medical Informatics and Decision Making.

    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 The authors say high heterogeneity among the included studies limits how the results can be interpreted.

    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