The concern is not hypothetical. AI systems learn from data, and when the data they scrape or are trained on reflect existing human stereotypes and healthcare disparities, the resulting models can form, copy, and amplify harmful patterns. In everyday consumer applications, biased AI is troubling enough, but in clinical settings the stakes rise sharply. An algorithm that performs well on average may quietly underperform for women, older patients, or members of ethnic minorities, and those failures can translate directly into missed diagnoses, delayed treatment, and worse outcomes. Because medical AI increasingly influences decisions about who gets scanned, who gets referred, and who gets treated, researchers argue that careful evaluation of potential demographic bias is not an optional refinement but a core safety requirement.

To understand how well the field is meeting that requirement, the Osaka Metropolitan University team conducted a scoping review, a form of systematic literature mapping designed to capture the breadth of research on a topic. Their goal was twofold: first, to identify studies that validated commercially available radiology AI products, and second, to document trends in how those studies reported results for demographic subgroups defined by sex, age, and ethnicity. Scoping reviews are particularly useful when a research landscape is sprawling and inconsistent, because they reveal not just what has been measured but what has been left out. In this case, what has been left out turned out to be the most important part of the safety picture.

The scale of the search was substantial. The team collected 545 studies covering 252 commercial AI products that included some reported demographic subgroup data. On its face, that number might suggest a healthy level of attention to demographics. But the picture changed dramatically when the researchers applied stricter criteria. Of those 545 studies, only 77, validating just 52 products, were found to include both demographic details and actual subgroup analysis results. In other words, fewer than one in seven of the studies they identified provided the kind of per-subgroup performance evidence needed to detect bias, and the validated products represented only a fraction of the commercial landscape.

The statistical dimension of the problem is where the findings become especially striking. The researchers applied the Wilson Confidence Interval formula, a standard method for calculating proportions and their uncertainty, to studies validating AI for tuberculosis detection. They found that 67 percent of the reported datasets were at risk of being underpowered for sex subgroup analysis. Underpowered means the dataset was too small to detect a meaningful difference in performance between subgroups even if one existed. An underpowered analysis can produce a reassuring result that is essentially meaningless, because the study simply lacked the statistical capacity to reveal a disparity. For clinicians relying on such studies, the absence of evidence of bias can be mistaken for evidence of the absence of bias.

Perhaps the most sobering conclusion is that this situation has not improved despite years of calls for better demographic reporting. The demand for improved reporting on sex, age, and ethnicity in medical AI has been voiced repeatedly by researchers, ethicists, and regulators, precisely because of the dangers that biased algorithms pose to marginalized groups. Yet the regression analysis used by the Osaka team to map trends across the collected studies revealed that performance reporting for demographic subgroups has not become more common over time. The field, in effect, has been accumulating validation studies without accumulating the evidence needed to confirm equitable performance, a pattern the authors describe as a systemic problem rather than a series of isolated oversights.

Dr. Walston characterized the landscape in stark terms. “This scoping review quantifies how fragmented the commercial validation landscape is, showing that reporting for both the demographics and per-subgroup performance is inadequate for estimating subgroup bias,” she stated. “This systemic problem requires effort from all stakeholders, from researchers to regulatory agencies, encouraging thorough reporting and commercial product validation to support physician and patient trust in medical AI products.” The statement frames the issue as one that no single actor can solve: journal reviewers, device manufacturers, hospital procurement teams, and government regulators all play a role in whether subgroup evidence is produced, demanded, and acted upon.

The fragmentation Dr. Walston describes has practical consequences for how medical AI reaches patients. When validation is scattered across many small studies, each examining a different product with different datasets and different reporting conventions, it becomes nearly impossible to compare products or to build a coherent picture of where bias risk concentrates. A hospital choosing between competing algorithms may find that one has published subgroup results and another has not, or that both report averages without revealing how performance varies by sex or ethnicity. Without standardized, transparent subgroup reporting, the market cannot reward fairness, and regulators cannot easily verify it. The result is a validation ecosystem that quantifies overall accuracy while leaving the distribution of that accuracy across patient groups largely unmeasured.

The study, published in European Radiology under the title describing the current state of demographic subgroup reporting for commercially available AI in radiology, arrives at a moment when health systems worldwide are scaling up their reliance on algorithmic tools. Tuberculosis detection, one of the applications examined in the power analysis, is a case in point: screening programs in high-burden countries increasingly depend on AI readers, and any systematic performance gap between demographic groups could quietly shape which communities receive timely diagnoses. The authors note that the findings expose the need for effective, transparent reporting to confirm the safe and unbiased performance of medical AI products across all patient subgroups, a requirement they argue underpins both clinical safety and public confidence.

The path forward, as the researchers frame it, involves collective accountability. Manufacturers would need to release demographic details and per-subgroup performance data for their products; researchers would need to design validation studies with sufficient statistical power to detect subgroup differences; and regulatory agencies would need to make such reporting a condition of approval rather than a voluntary courtesy. The authors declare a relationship with the Medical AI Promotion Institute, Inc., in their conflict of interest statement, underscoring how closely the commercial and academic sides of this field are intertwined. What the review makes clear is that trust in medical AI cannot be assumed from headline accuracy figures alone. Until subgroup evidence becomes routine, the possibility that these tools are quietly copying and amplifying healthcare bias remains an open question, one that patients, physicians, and regulators deserve to have answered with data rather than assurances.

Subject of Research: Demographic subgroup reporting in validation studies of commercially available radiology AI products

Article Title: Check for copycat bias in medical AI

Article References: Check for copycat bias in medical AI. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: medical AI, radiology, algorithmic bias, demographic subgroups, scoping review, Osaka Metropolitan University, European Radiology, tuberculosis detection, validation studies, health equity, regulatory reporting, statistical power

At the heart of the debate lies a conceptual tension that will be familiar to anyone who studies mental health measurement. The Impostor Phenomenon, the persistent feeling that one’s accomplishments are undeserved and that one is a fraud despite evidence of competence, is not a diagnosis in the way that, say, major depressive disorder is. It is conceptually a continuous construct, distributed along a spectrum across the population. Critics argue that applying the epidemiological language of prevalence to such a construct is conceptually incorrect, because prevalence traditionally refers to the proportion of a population that has a categorical condition. You either have an infection or you do not; you cannot, in the same strict sense, have a fraction of impostorism.

Salari and colleagues do not dispute this premise. In their response, they explicitly agree that the Impostor Phenomenon is continuous and that referring to its prevalence in the strict epidemiological sense may be conceptually wrong. But they argue that this does not rule out threshold-based prevalence estimates, and they point to well-established precedents. Depression and anxiety are also conceptualized as continua, yet researchers routinely estimate their prevalence using instruments such as the PHQ-9 for depression and the GAD-7 for anxiety, which employ empirically derived thresholds to classify individuals above a cutoff. In those fields, decades of validation work have established which scores correspond to clinically meaningful impairment, giving the thresholds legitimacy. The question the exchange raises is whether the Impostor Phenomenon field has earned the same privilege.

The instrument in question is the Clance Impostor Phenomenon Scale, or CIPS, a questionnaire that yields a continuous score of impostor feelings. Primary studies included in the meta-analysis reported the percentage of health service providers scoring above various cutoffs on this scale, effectively equating the proportion of high scorers with prevalence. Salari’s team then pooled those reported proportions into a single estimate. Here the authors make a striking concession: they fully acknowledge that the current cutoffs for the CIPS are inconsistently applied across the existing literature and lack systematic validation. Different studies have used different thresholds, sometimes without empirical justification, meaning that two studies of identical populations could report wildly different prevalence figures depending solely on where they drew the line.

This is precisely the point Brauer and Proyer drove home with an empirical demonstration. By applying six different cutoffs to the same datasets, they showed that the resulting proportions of cases diverged dramatically. They also argued that the extremely high heterogeneity in the meta-analysis, quantified by an I-squared statistic of 98.6 percent, is evidence that pooling prevalence rates in the first place is invalid. In meta-analysis, I-squared describes the percentage of variation across studies that is due to real differences rather than chance; a value near 100 percent means the studies are barely measuring the same thing. To the critics, pooling estimates from studies that use incompatible thresholds produces a number that means little.

Salari and colleagues offer a different reading of that statistic. In their view, the high heterogeneity is itself an important empirical finding, a quantitative portrait of a literature that has operationalized its central construct inconsistently. Meta-analysis, they argue, is precisely the tool that allows researchers to measure and expose this problem, highlighting how different thresholds, samples, and instruments contribute to variance. Without such a synthesis, the field would lack any systematic appreciation of how widely prevalence estimates diverge. They also note that some of the cutoffs applied in the critics’ demonstration were arbitrary and drawn from two convenience datasets that are not population-representative, which they contend artificially inflates the apparent instability and should not be generalized to broader populations. Demonstrating variability across arbitrary thresholds, they write, does not invalidate prevalence research; it underscores the need for consensus and validation studies.

The exchange also touches on a provocative interpretive question. Brauer and Proyer suggested that if 62 to 70 percent of individuals report impostor feelings, as some estimates imply, then the experience should perhaps be regarded as normative rather than pathological. The Salari team counters that high prevalence does not imply triviality. They draw an analogy to physician burnout, which large systematic reviews have shown to be reported by a strikingly high proportion of doctors, yet which has clear and consequential implications for health care delivery. Similarly, they argue, impostor feelings, even if common, are associated with impaired well-being, reduced job satisfaction, and elevated burnout risk. Prevalence estimates, on this view, remain useful for scoping the potential consequences of a widespread experience, regardless of whether it is labeled a disorder.

Notably, the authors point out that their original review contained findings that survive the critique largely intact. Beyond the contested prevalence figures, the meta-analysis summarized correlations between impostor feelings and constructs such as self-esteem, anxiety, depression, stress, and burnout. Because those analyses used continuous scores rather than categorical cutoffs, they are conceptually reliable and less dependent on the threshold decisions that lie at the center of the dispute. This distinction matters for the field’s future direction: the relational findings, which describe how impostorism moves alongside other psychological variables, rest on firmer measurement ground than the headline prevalence numbers.

The response also ventures into a broader reflection on the nature of psychological measurement itself. The authors observe that screening tools for mental health conditions inevitably involve trade-offs, with all instruments producing false positives and false negatives, and that different thresholds can be deployed to compensate for these limitations and optimize screening performance. A screening tool, they emphasize, is not a diagnosis; it flags individuals for further clinical attention. Their stated goals in examining the Impostor Phenomenon were twofold: to draw attention to a condition that has received comparatively little scrutiny, and to provide clinicians with usable information. On that view, the very act of publishing a contested prevalence estimate, and engaging publicly with its critics, serves the scientific process by forcing the field to confront how much its conclusions depend on arbitrary measurement decisions.

Where does this leave the science of feeling like a fraud? The two sides converge more than the heated framing might suggest. Salari’s team agrees that any prevalence estimates should be presented with caution and explicit caveats until robust validation studies are conducted, and they endorse the critics’ implicit agenda: large-scale validation of CIPS thresholds against external criteria, studies using population-representative samples rather than convenience datasets, and community-wide consensus on cutoff standards, modeled on the practices that stabilized depression and anxiety research. The meta-analysis, they argue, does not endorse any single cutoff but documents the consequences of their inconsistency, functioning as a warning, while the commentary supplies the conceptual framework for understanding that warning. Both papers, read together, point toward the same next steps. For the millions of health care workers who quietly doubt their own competence, the practical takeaway is unchanged: the feelings are real, common, and linked to burnout. But for researchers hoping to put a single trustworthy number on how many people experience them, the message from this exchange is clear: the field first needs to agree on where the line should be drawn, and then prove that the line means something.

Subject of Research: Methodological debate over estimating prevalence rates of the Impostor Phenomenon using the Clance Impostor Phenomenon Scale

Article Title: Response to: Matters Arising from Salari et al. (2025): Why it is (currently) impossible to estimate accurate prevalence rates of the Impostor Phenomenon

Article References: Salari, N., Hashemian, S. H., Hosseinian-Far, A., Fallahi, A., Heidarian, P., Rasoulpoor, S., & Mohammadi, M. (2026). Response to: Matters Arising from Salari et al. (2025): Why it is (currently) impossible to estimate accurate prevalence rates of the Impostor Phenomenon. BMC Psychology, 14(1), Article 1400. https://doi.org/10.1186/s40359-026-05555-6

Image Credits: AI Generated

DOI: 10.1186/s40359-026-05555-6

Keywords: Impostor Phenomenon, prevalence, meta-analysis, Clance Impostor Phenomenon Scale, CIPS cutoffs, heterogeneity, psychological measurement, health service providers, burnout, screening thresholds, BMC Psychology, validation studies

The new analysis, published as a systematic review and meta-analysis in Food Science & Nutrition, was conducted by a team at the University of Catanzaro in Italy. The researchers combed through three major bibliographic databases, PubMed, Web of Science, and Scopus, covering publications from January 2000 through February 2026. From 5,247 initially identified papers, they narrowed the field through duplicate removal, title and abstract screening, and full-text assessment to 43 validation studies encompassing 8,467 healthy adults. Every included study had tested a food frequency questionnaire against an established reference method, either repeated 24-hour dietary recalls or prospective food records, and reported Pearson or Spearman correlation coefficients for nutrient intake.

The restrictions were deliberate and methodologically significant. The team limited inclusion to studies conducted in Europe, the United States, Canada, Australia, and New Zealand, regions that share mature, harmonized food composition databases such as the USDA nutrient tables, the UK’s McCance and Widdowson compilations, and the EuroFIR and FSANZ systems. They also excluded studies with fewer than 50 participants, questionnaires designed for single nutrients, and populations with specific diseases or physiological states such as pregnancy. The goal was not to assume that people in these countries eat the same food, but to ensure that the measurement infrastructure behind each validation study was broadly comparable, reducing a major source of statistical noise before the numbers were ever pooled.

The pooled results tell a story of moderate but meaningful validity. Against 24-hour recalls, the correlation coefficient for total energy was 0.454, rising to 0.501 when food records served as the reference. Carbohydrates fared better at roughly 0.52 against both comparators, and dietary fiber emerged as the standout performer, with correlations of 0.593 against recalls and 0.598 against records. Protein and lipids landed in the 0.41 to 0.48 range. Among micronutrients, riboflavin and magnesium showed some of the strongest associations, reaching 0.633 and 0.573 respectively against food records, while vitamin C, calcium, and zinc performed solidly. At the other end of the spectrum, iodine was the clear weak point, with a correlation of just 0.103 against recalls, and vitamin A and sodium also lagged behind.

Why does fiber shine while iodine stumbles? The answer lies in the psychology and physiology of dietary recall. Fiber comes bundled with foods that people recognize and count easily: whole grains, legumes, fruits, vegetables. Iodine, by contrast, hides in iodized salt, dairy, and seafood in amounts that vary enormously and that few respondents can estimate. Sodium poses a similar problem, since discretionary salt added at the table or during cooking is nearly invisible to any self-report instrument. The researchers also noted that monounsaturated and polyunsaturated fats correlated more weakly than saturated fat, likely because participants struggle to identify the types of oils and spreads they use, a well-documented blind spot in self-reported diet.

Correlation coefficients, however, only capture one dimension of questionnaire quality: the ability to rank individuals correctly. The team therefore also calculated standardized mean differences to detect systematic bias in absolute intake levels. Here a familiar pattern emerged. Compared with 24-hour recalls, questionnaires tended to overestimate the intake of several nutrients, with significantly elevated differences for fiber, potassium, vitamin E, and polyunsaturated fats. The overestimation was smaller when food records were the reference, and for many nutrients the differences were not statistically significant. This bias likely reflects a combination of difficulty in judging average portion sizes and a well-known tendency to overreport foods perceived as healthy. Crucially, the authors note that such bias can be partially corrected statistically, and that for most nutrients the standardized mean differences remained below 0.50, a threshold suggesting acceptable agreement.

One of the most valuable contributions of the analysis is its set of subgroup comparisons, which function as a practical design guide for the next generation of questionnaires. Interviewer-administered versions outperformed self-administered ones for energy, carbohydrates, protein, and lipids, presumably because a trained interviewer reduces misinterpretation of questions and improves portion-size estimation. Questionnaires with more food items, above the median of 128 in this dataset, showed slightly higher correlations than shorter instruments, consistent with the idea that a longer food list better captures the true variability of habitual diets, though at the cost of greater respondent burden. Studies rated as very good on the Serra-Majem quality scale reported stronger agreement, and European studies edged out those from other regions. Notably, sample size above the recommended minimum of about 50 participants did not materially change validity, although larger studies displayed far greater heterogeneity.

That heterogeneity is the elephant in the room. The I-squared statistic, which quantifies the share of variability across studies that cannot be explained by chance, exceeded 75 percent for most nutrients and approached or surpassed 90 percent for several, including energy and fiber. Even after restricting the analysis to methodologically homogeneous Western countries, the inconsistency persisted, leading the authors to conclude that instrument-level factors, questionnaire length, administration mode, reference period, and nutrient-specific measurement error, drive variability more than cultural or population differences do. This finding echoes the largest previous meta-analysis of food frequency questionnaire validity, published by Cui and colleagues in 2023, which pooled 130 studies and 21,494 participants globally and reported validity coefficients ranging from roughly 0.17 to 0.77. The new study extends that work by adding Scopus as a third database, pushing the search window six years forward to capture the growing wave of digital and web-based questionnaires, and sharpening the geographic focus for interpretability in Western epidemiological contexts.

The practical implications are considerable. For researchers designing cohort studies, the message is that overall-diet questionnaires are well suited to their classic epidemiological role: ranking individuals by habitual intake of energy, macronutrients, carbohydrates, and fiber, and characterizing broad dietary patterns. They are a weaker choice when the exposure of interest is a specific micronutrient with high day-to-day variability or hidden dietary sources, such as iodine or sodium; in those cases, targeted validation, repeated recalls, or objective biomarkers remain essential. The authors also caution that their findings concern relative validity, the ability to rank people, rather than absolute agreement, since the eligibility criteria required correlation coefficients and excluded studies reporting only complementary measures such as Bland-Altman analyses.

Looking forward, the transition from paper to digital formats looms large. Web-based questionnaires promise standardized administration, integrated portion-size images, automated nutrient calculation, and better acceptability among digitally fluent participants, and several such tools appeared among the studies included here. Yet the authors are clear-eyed about the limits of technology: systematic underreporting and memory bias are intrinsic to self-report, and digitization alone cannot eliminate them. Their conclusion is a measured endorsement. The food frequency questionnaire remains irreplaceable for large-scale nutrition research because of its low cost, low burden, and unique ability to capture long-term habitual intake, but its results should be read with an understanding of what it measures well and what it does not. Future work, they argue, should focus on identifying the conditions under which these instruments deliver the most informative estimates, so that the humble checklist continues to earn its central place in the science of diet and disease.

Subject of Research: Relative validity of food frequency questionnaires for assessing dietary intake in adult populations

Article Title: Relative Validity of Food Frequency Questionnaires in Adult Populations: A Systematic Review and Meta‐Analysis

Article References: Pujia, C., Ferro, Y., Placanica, A., Bruno, R., Guadagni, F., Malaguarnera, G., Mazza, E., Montalcini, T., & Pujia, A. (2026). Relative Validity of Food Frequency Questionnaires in Adult Populations: A Systematic Review and Meta‐Analysis. Food Science & Nutrition, 14(10), Article e72381. https://doi.org/10.1002/fsn3.72381

Image Credits: AI Generated

DOI: 10.1002/fsn3.72381

Keywords: food frequency questionnaire, dietary assessment, nutritional epidemiology, meta-analysis, systematic review, validation studies, 24-hour recall, food records, macronutrients, micronutrients, measurement error, diet-disease risk