Canada

Artificial Intelligence (AI) is a tool that is increasingly being used in dermatology. However, systems are not always producing the same results in differing skin tones. Inequitable performance between groups can worsen health inequities. This study analyzes whether AI diagnostic perfor- mances differ between Fitzpatrick skin type groups for classifying malignant and non-malignant skin conditions using the deep-learning model MobileNetV2. The data set used was Fitzpatrick17k, which includes 16,577 pictures of skin conditions, labeled by diagnosis and Fitzpatrick skin tone classification (I-VI). Performance was measured through accuracy, precision, recall, and subgroup comparison between skin tone groups. The Fitzpatrick17k dataset was found to be unbalanced with a 2.2 times higher rates of lighter skin tone images (Fitzpatrick I-III) than darker skin tone images (Fitzpatrick IV-VI). Sensitivity analysis across multiple thresholds showed that recall rates were also consistently lower in IV-VI images than in I-III images. These results suggest that the imbalance of data sets can contribute to unfair diagnostic performance and emphasize the need for more representative dermatology datasets.

Equitable Health Technologies: Overcoming Biases and False Inferences in AI Diagnostics and Drug Discovery

Assessment of Diagnostic Performance Across Fitzpatrick Skin

skin diseases

fitzpatrick skin types

fitzpatrick17k

convolutional neural network

artificial intelligence

As AI transforms healthcare, ensuring reliability is critical. However, systems learning from existing data often reflect gaps or inconsistencies, leading to false inferences, diagnostic errors, and unequal performance across patient groups.

Open Data is pivotal for transparency and accountability. Biases emerging from unrepresentative datasets or flawed assumptions can perpetuate health disparities.

Leveraging public data on genomics, clinical trials, and social determinants allows researchers to audit and mitigate these risks effectively.

We task undergraduate and graduate students to harness data science to:

* Detect, reduce, and prevent bias in diagnostics and drug discovery;

* Develop data-driven tools that increase fairness and inclusivity;

* Innovate bias-aware model evaluation techniques to build trustworthy AI systems.

As AI transforms healthcare, ensuring reliability is critical. However, systems learning from existing data often reflect gaps or inconsistencies, leading to false inferences, diagnostic errors, and unequal performance across patient groups.

technical paper

Artificial intelligence is increasingly used to predict new drug treatments by learning from large biomedical databases. If those databases describe some diseases far better than others, AI predictions may inherit that imbalance while appearing equally reliable for all. We tested whether this occurs for neglected tropical diseases (NTDs), a group of conditions that affect over a billion people in low-income regions yet receive minimal research investment. Using Open Targets, a widely used drug–disease evidence platform, we trained a knowledge graph model on 759,217 evidence links and audited its predictions across 96 diseases, introducing the clinical translation ratio: the share of a disease's evidence drawn from clinically actionable sources rather than basic research. NTDs had a 42-fold lower clinical translation ratio than other diseases (0.5% versus 20.4%, p = 3.49 × 10⁻⁸). The gap held across four evidence-classification schemes, persisted against matched infectious diseases, and widened over three database releases. Twelve of nineteen NTDs had no clinically actionable evidence in any release. Most concerning, the model did not account for this scarcity. It was equally confident in its NTD predictions despite near-zero supporting evidence, and its top NTD predictions were three times less likely to correspond to known treatments. The model inherits the evidence gap and hides it behind uniform confidence. To address this, we built an open evidence-equity auditing framework and dashboard that reveals when AI confidence rests on evidence poverty. Before AI can accelerate drug discovery, it must first show whose evidence it is missing.

AI Drug Discovery Inherits Evidence Inequity for Neglected Tropical Diseases

There has been an increase in the usage of machine-learning models to analyze cancer transcriptomic data. However, imbalanced datasets can make models appear reliable while underperforming on underrepresented cell populations. This manuscript investigated false inferences when classifying malignant versus non-malignant tumor microenvironment (TME) observations based on class imbalance in a breast cancer Slide-seqV2 spatial transcriptomic dataset. An .h5ad dataset containing 5,833 observations and 12,630 gene-expression features was analyzed in Python using Google Colab. Binary labels were created from cell-type annotations with malignant and non-malignant TME. Results revealed the dataset was highly imbalanced, with malignant observations representing 94.77% of observations. After preprocessing raw counts from adata.raw.X, principal component analysis, a majority-baseline model, logistic regression, confusion matrices, balanced accuracy, minority-class recall, and feature-importance analysis were performed to evaluate model behavior. The majority-class baseline achieved 94.79% accuracy but failed to identify any non-malignant TME observations. Logistic regression achieved 91.30% accuracy and improved balance accuracy to 64.32% but still misclassified 50 of 76 non-malignant TME observations as malignant. Thus, although the majority-class baseline achieved higher overall accuracy, it provided no minority-class detection, whereas logistic regression showed improved performance across both classes. These results show that overall, accuracy can create misleading confidence when evaluating classification performance in this highly imbalanced spatial transcriptomic dataset. The findings illustrate the importance of using imbalance-sensitive evaluation metrics when assessing machine-learning performance in spatial transcriptomic classification.

Class Imbalance and False Inference in Breast Cancer Spatial Transcriptomic Classification

Overlap-based drug and compound prioritization pipelines are attractive because they are scalable, transparent, and easy to apply across large drug-disease search spaces. However, they are also vulnerable to annotation-breadth effects, whereby broadly annotated genes can inflate apparent support across many disease contexts. We developed a bias-aware computational framework for candidate prioritization in Alzheimer disease, Parkinson disease, and amyotrophic lateral sclerosis (ALS). The framework combines a normalized raw overlap score, annotation-breadth correction using gene-specificity weights, and a matched-null permutation layer applied to a selected high-priority subset. Across 62,887 overlap-positive pairs, correction substantially changed rank order and reduced the contribution of broadly annotated genes. However, exploratory downstream alignment with ClinicalTrials.gov remained weak, and a small curated therapeutic-history benchmark showed mixed recovery of known positives. These findings support the framework as a transparent hypothesis-generation and triage tool, but not as a predictor of clinical efficacy.

A Bias-Aware Computational Framework for Drug and Compound Prioritization across Neurodegenerative Diseases

Artificial intelligence (AI) frameworks combining 3D Computed Tomography (CT) scans with patient metadata hold promise for advancing lung cancer staging. However, these systems face vulnerabilities from shortcut learning, where models exploit superficial data correlations rather than authentic biological features. This study evaluates and quantifies the structural and optimization risks of metadata-driven algorithmic bias during the multi-class classification of primary lung tumor stages (T1 to T4). Utilizing a multi-institutional dataset of 354 patients, we constructed and compared two parallel neural network frameworks: a unimodal 3D EfficientNet-B3 image-only baseline and a multimodal late-fusion network (FusioNet-B3) that combined 3D chest CT tensors with tabular clinical metadata, including smoking history. Across experimental training runs, both architectures failed to generalize beyond the dataset’s majority class distribution (~42% validation accuracy). Notably, under an accelerated learning rate (η = 1x10-4), the multimodal network experienced complete feature suppression. It bypassed high-dimensional visual anatomy entirely, utilizing the binary smoking history token as an algorithmic shortcut to minimize mathematical loss. This resulted in a zero-variance, frozen validation accuracy flatline for 199 consecutive epochs. These findings provide empirical proof that naive integration of highly correlated clinical priors creates competitive optimization dynamics that can render image-processing branches obsolete. This failure underscores that small cohort sizes and severe class imbalances force networks to rely on demographic biases rather than genuine oncological features. To ensure robust, safe, and equitable clinical deployment, future multimodal architectures must shift away from basic concatenation toward explicit debiasing frameworks like adversarial training to force model grounding in true visual pathology.

Shortcut Learning Risks in Multimodal Deep Learning for Lung Cancer Staging

Clinical artificial intelligence models trained on imbalanced health data systematically underperform for minority subpopulations, raising concerns about whether synthetic data genuinely improve fairness or merely hide underlying bias. This study investigates whether synthetic data augmentation via the Synthetic Minority Oversampling Technique (SMOTE) and Bayesian hierarchical modelling produce meaningful fairness improvements in a diabetes prediction task using the National Health and Nutrition Examination Survey (NHANES) 2019–2020 and 2021–2023 cycles. Three models are developed: a logistic regression baseline trained on imbalanced data, a SMOTE- augmented logistic regression, and a Bayesian hierarchical GLM with per-race intercepts estimated via MCMC sampling. All models are assessed on AUC, F1, precision, and recall stratified by race/ethnicity and sex, with SHAP explainability analyses applied to both logistic regression models and posterior convergence diagnostics reported for the Bayesian model. The baseline model achieves an overall AUC of 0.897 but a recall of 0.524. SMOTE increases recall across all subgroups (overall recall 0.749) but produces asymmetric precision losses and fails to improve AUC, suggesting the model is more willing to predict the positive class, rather than being more accurate. The Bayesian hierarchical model achieves the most balanced precision- recall tradeoff for several subgroups, while posterior diagnostics confirm well-converged chains and modest but non-negligible between-group variance (sigma_alpha ≈ 0.2). This suggests that SMOTE improves sensitivity at the cost of precision without reducing subgroup discrimination gaps, while Bayesian hierarchical modelling helps by estimating uncertainty and capturing explicit group-level differences. Subgroup- stratified evaluation is essential for honest clinical AI assessment.

Does Synthetic Data Fix or Hide Bias?

Breast cancer detection relies heavily on mammographic imaging, but dense tissue can make malignant lesions harder to identify as both dense tissue and tumours appear radiographically similar. With increasing use of AI-based diagnostic tools in medical imaging, it is important to determine whether these models are affected by the same limitations that challenge human interpretation. The purpose of this research is to emphasize the need to train AI models on a more representative dataset that includes high-noise samples to ensure equitable care across demographics. This study investigates whether variability in breast tissue density acts as a confounding factor in AI-driven tumour detection thus leading to diagnostic inaccuracies. Mammographic images from the CBIS-DDSM dataset were analyzed using Mammo-CLIP, a pretrained model for mammography-based classification. Cases were grouped by BI-RADS breast density categories, and model performance was evaluated using sensitivity, specificity, accuracy, FNR, and chi-squared testing between fatty and dense breast tissue groups. Model sensitivity decreased from 64% to 34% and FNR increased from 36.1% to 66.1% with increasing breast density. Classification accuracy was significantly higher in fatty tissue than dense tissue, and cancer detection amongst malignant cases was also significantly reduced in dense tissue. These findings suggest that denser tissue negatively influences Mammo-CLIP’s ability to detect malignant lesions and increases likelihood of missed cancers. This has important clinical implications, as false negatives may delay diagnosis and treatment, particularly for patients with dense breast tissue.

Effect of breast tissue density variability on AI-based tumor detection

Healthcare systems generate vast quantities of clinical data, yet much of this information remains locked within unstructured free-text records, including physician notes and doctor-patient transcripts, making it inaccessible for large-scale analysis. Manually extracting structured information from this data is time-consuming, expensive, and error-prone. Large language models (LLMs) have demonstrated promising performance in a variety of medical fields, including radiology report summary, doctor-patient dialogue transcription, clinical question answering, and diagnostic reasoning. This study evaluates four LLMs (GPT-5.4 mini, Qwen3 32B, Llama 3.1 8B, and Gemini 3.1 Flash-Lite Preview) on their ability to label radiology diagnostic conclusions and clinical assessment summaries from unstructured medical text using MedVAL-Bench, a 2025 Stanford MIMI benchmark with physician-validated reference outputs. The models were tested on two tasks: doctor-patient dialogue summarization (dialogue2note, n=85) and radiology report impression extraction (report2impression, n=190), which were assessed using ROUGE-1, BERTScore F1, and BLEU. The results suggest that proprietary models like GPT-5.4 mini outperformed all models with a ROUGE-1 of 0.376 and BERTScore F1 of 0.860, while Qwen3 32B, Llama 3.1 8B, and Gemini 3.1 Flash-Lite Preview produced comparable results. ROUGE-1 and BLEU scores were consistently low across all models, reflecting the limitations of n-gram metrics in clinical contexts, where a single-word difference can carry critical medical significance. BERTScore F1 emerged as the most reliable indicator of semantic similarity. These findings suggest LLMs can significantly automate clinical text labelling, offering a scalable, low-cost alternative to manual data extraction in healthcare research.

A Comparative Analysis of LLM-Assisted Labelling of Unstructured Clinical Data

General-purpose artificial intelligence chatbots are increasingly used for health information, despite not being validated as medical devices. In 2026, ECRI ranked misuse of AI chatbots as the leading health technology hazard, noting that human-sounding responses can mislead users when information affects patient care. Dermatology is a particularly important domain for AI safety because patients can easily photograph visible skin concerns and seek advice before entering formal care. Although specialized dermatology models such as SkinGPT-4 demonstrate the feasibility of image-based AI analysis, widely available general-purpose tools such as ChatGPT, Claude, and Gemini may be more accessible to ordinary users. This paper proposes SAFER-Skin, a reproducible safety-audit framework for evaluating the off-label use of general-purpose multimodal chatbots when responding to consumer-style dermatology image queries. Rather than diagnostic accuracy, the framework is designed to evaluate safety behavior: whether chatbots provide appropriate referral advice, communicate uncertainty, avoid false reassurance, and refrain from prioritizing skincare products when medical evaluation may be needed. The proposed methodology uses publicly available dermatology image datasets including the Diverse Dermatology Images dataset and the Skin Condition Image Network, paired with neutral, reassurance/product-seeking, and safety-first prompt templates. Responses would be coded using a structured rubric with a primary safety classification and seven binary safety features, then stratified by model, prompt style, risk tier, dataset, and available skin-tone metadata. No empirical results are reported; this manuscript presents the proposed framework and protocol. SAFER-Skin is intended for AI safety researchers, health AI developers, and clinical governance teams. It does not validate AI for clinical diagnosis or replace dermatologists; it provides a practical method for identifying unsafe response patterns before off-label AI use in dermatology becomes entrenched.

SAFER-Skin: Auditing AI Skin Advice Across Skin Tones

Traumatic brain injury (TBI) is an important public health concern in Canada and is frequently associated with depression – although often underdiagnosed in this population. Large-scale population health datasets, like ones provided by Statistics Canada, typically rely on clinical diagnoses as outcome variables. As these datasets are increasingly used to develop artificial intelligence (AI) diagnostic models, systematic disparities in diagnosis may become embedded within algorithmic systems – contributing to existing health inequities. This study examined the relationship between TBI and diagnosed mood disorders in Canada – exploring demographic differences that may influence diagnosis and healthcare representation. Data were obtained from the Canadian Community Health Survey (CCHS) 2017–2018 Public Use Microdata File. Descriptive analyses, chi-square tests, and multivariable logistic regression compared mood disorder prevalence between respondents reporting TBI and those reporting other injuries, controlling for sex, age, and immigrant status. The final sample included 15,884 respondents, of whom 611 reported a TBI. Mood disorders were more prevalent among TBI survivors (18.6%) than among those reporting other injuries (14.3%). TBI remained significantly associated with increased odds of reporting a mood disorder (OR = 1.33, p = .008) after adjustment for demographic variables. Women with TBI reported higher rates of mood disorders than men (22.8% vs. 13.3%), and non-immigrants demonstrated greater odds of diagnosis than immigrants. These findings suggest that depression remains a significant consequence of TBI, highlighting how demographic differences in diagnosis may influence datasets used to develop AI technologies; underscoring the importance of equitable data representation in future healthcare innovation.

Bias in Depression Diagnosis Following Traumatic Brain Injury: Implications for Equitable AI Diagnostics in Canada

Clinical trials provide the evidence used to guide treatment recommendations, regulatory decisions, and healthcare practice. As clinical evidence increasingly supports precision medicine and AI-assisted healthcare, the completeness and representativeness of clinical trial data have become important considerations. This study examined demographic representation and reporting practices across cardiovascular, respiratory, and diabetes/metabolic clinical trials identified through ClinicalTrials.gov. A cross-sectional analysis of 54 completed studies was conducted, including 18 cardiovascular studies, 18 respiratory studies, and 18 diabetes/metabolic studies. Data were collected on sex distribution, race and ethnicity reporting, enrollment, and study design. Descriptive analyses compared demographic representation and reporting across disease areas. Differences were observed across all three disease categories. Cardiovascular studies contained the most missing demographic information, with 11,997 participants lacking sex-specific reporting and 11,088 participants classified as Unknown/Not Specified for race or ethnicity. Respiratory studies generally demonstrated more complete demographic reporting, while diabetes/metabolic studies reported the highest number of female participants. Interventional studies comprised 45 of the 54 included studies. These findings suggest that demographic representation and reporting practices remain inconsistent across clinical research. Incomplete demographic reporting may contribute to evidence gaps that limit the interpretation and generalizability of clinical findings. As clinical trial datasets increasingly inform precision medicine, predictive analytics, and AI-assisted healthcare systems, improving demographic inclusion and reporting standards may strengthen the quality, reliability, and applicability of future healthcare evidence and technologies.

Downloads

Next from Equitable Health Technologies: Overcoming Biases and False Inferences in AI Diagnostics and Drug Discovery

AI Drug Discovery Inherits Evidence Inequity for Neglected Tropical Diseases

Similar lecture

Skin-Tone Fairness in Dermatology AI: Limits of Pretraining and Harmonization

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from Equitable Health Technologies: Overcoming Biases and False Inferences in AI Diagnostics and Drug Discovery

AI Drug Discovery Inherits Evidence Inequity for Neglected Tropical Diseases

Similar lecture

Skin-Tone Fairness in Dermatology AI: Limits of Pretraining and Harmonization

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads