Zhaohui Liang

Background: Despite high reported accuracy on clinical and evidence appraisal tasks, AI-generated medical information may lack explicit support from source documents. This creates challenges for digital health practitioners regarding transparency, auditability, and trust when AI systems are used for evidence synthesis, guideline development, and clinical knowledge management. Large language models (LLMs) can generate fluent and seemingly correct outputs, but existing evaluations often rely on agreement with human judgments and do not directly assess whether AI-generated content is grounded in underlying evidence. Objective: This study measures evidence support and hallucination in AI-generated medical information by assessing the extent to which LLM-generated risk-of-bias assessments are supported by source clinical trial reports. Methods: We evaluated 3 LLMs (GPT-5, OpenAI o3-mini, and GPT-3.5) on risk-of-bias (RoB 2) assessment using all 97 randomized controlled trials for which the source Cochrane systematic review provided complete human RoB 2 annotations and full-text reports were accessible, constituting the complete reference set. No train-validation split was applied; all 97 studies were used for evaluation. Model outputs were constrained to structured RoB 2 signaling questions and domain-level judgments. For each generated claim, relevant text passages were retrieved from trial reports using the Okapi BM25 (Best Matching 25) algorithm. A verification step assigned evidence verdicts (supported, contradicted, not found, or out of scope) with verbatim quotations. We quantified evidence support rates and conservative and strict hallucination rates. Task performance was evaluated using exact and binary accuracy, sensitivity, specificity, F1-score, Youden J, and agreement with human reviewers using Cohen κ and Fleiss κ. Results: Binary accuracy of AI-generated risk-of-bias judgments was high across domains (90%-98%), whereas exact accuracy was substantially lower (42%-71%), reflecting frequent disagreements in severity classification despite correct directional classification. GPT-5 achieved the strongest overall performance, including perfect binary accuracy for overall risk-of-bias conclusions and the highest agreement with human reviewers (quadratic κ up to 0.81). However, evidence support rates across models ranged from only 60% to 65%, with conservative hallucination rates of 34%-37%. GPT-5 showed the highest mean evidence support (64.3%) and the lowest strict hallucination rate (35.7%). Mean top-1 BM25 retrieval scores were similar across models (approximately 30-31), suggesting that differences in hallucination were not primarily attributable to differences in retrieval strength. Conclusions: AI-generated medical information can achieve high decision-level accuracy while still lacking documentary support in a substantial proportion of outputs. Measuring evidence support and hallucination reveals important limitations that are not captured by agreement metrics alone. Retrieval-based evidence verification provides a reproducible and transparent approach for evaluating the reliability of AI-generated medical information, with direct relevance to digital health practice, evidence-based medicine, and medical informatics.

Multimodal learning using images and associated clinical text offers richer semantic supervision for medical AI. However, models trained with synthetic reports risk hallucination, and conventional multimodal tuberculosis (TB) systems are impractical because they require text at inference. In realworld screening workflows, particularly in low-resource settings or during triage, radiology reports are often unavailable or delayed. Computer-aided detection systems for chest X-rays (CXRs) are considered a potential solution. In this context, this study proposes a method that uses clinically grounded text as privileged information during training to improve a binary CXR classifier, while enabling image-only TB prediction at deployment. Frontal CXRs from Shenzhen (internal train/validation/test), Montgomery County, TBX11K, and NIAID TB Portals (external tests) were lung-cropped using a YOLOv8s detector and resized to 224 × 224. For Shenzhen, de-identified metadata and brief clinical notes were converted into structured reports encoding population type, TB status, laterality, lobar involvement, and adjunct findings; a parallel model used raw notes. A VGG-11 vision encoder and frozen CXR BERT text encoder were co-trained in a shared 256-dimensional space using image classification, cosine similarity, and supervised contrastive alignment losses. At inference, the text branch was removed, yielding an image only classifier regularized through multimodal supervision. Multimodal training with report supervision consistently improved image-only predictions, with structured report outperforming raw notes. Across internal and external cohorts, performance gains were reflected in higher balanced accuracy, Matthews correlation coefficient, and area under the curve. UMAP embeddings showed clearer class separation, and Grad CAM maps demonstrated improved localization of TB-relevant lesions.

BACKGROUND Despite high reported accuracy on clinical and evidence appraisal tasks, AI-generated medical information may lack explicit support from source documents. This creates challenges for digital health practitioners regarding transparency, auditability, and trust when AI systems are used for evidence synthesis, guideline development, and clinical knowledge management. Large language models (LLMs) can generate fluent and seemingly correct outputs, but existing evaluations often rely on agreement with human judgments and do not directly assess whether AI-generated content is grounded in underlying evidence. OBJECTIVE This study measures evidence support and hallucination in AI-generated medical information by assessing the extent to which LLM-generated risk-of-bias assessments are supported by source clinical trial reports. METHODS We evaluated 3 LLMs (GPT-5, OpenAI o3-mini, and GPT-3.5) on risk-of-bias (RoB 2) assessment using all 97 randomized controlled trials for which the source Cochrane systematic review provided complete human RoB 2 annotations and full-text reports were accessible, constituting the complete reference set. No train-validation split was applied; all 97 studies were used for evaluation. Model outputs were constrained to structured RoB 2 signaling questions and domain-level judgments. For each generated claim, relevant text passages were retrieved from trial reports using the Okapi BM25 (Best Matching 25) algorithm. A verification step assigned evidence verdicts (supported, contradicted, not found, or out of scope) with verbatim quotations. We quantified evidence support rates and conservative and strict hallucination rates. Task performance was evaluated using exact and binary accuracy, sensitivity, specificity, F1-score, Youden J, and agreement with human reviewers using Cohen κ and Fleiss κ. RESULTS Binary accuracy of AI-generated risk-of-bias judgments was high across domains (90%-98%), whereas exact accuracy was substantially lower (42%-71%), reflecting frequent disagreements in severity classification despite correct directional classification. GPT-5 achieved the strongest overall performance, including perfect binary accuracy for overall risk-of-bias conclusions and the highest agreement with human reviewers (quadratic κ up to 0.81). However, evidence support rates across models ranged from only 60% to 65%, with conservative hallucination rates of 34%-37%. GPT-5 showed the highest mean evidence support (64.3%) and the lowest strict hallucination rate (35.7%). Mean top-1 BM25 retrieval scores were similar across models (approximately 30-31), suggesting that differences in hallucination were not primarily attributable to differences in retrieval strength. CONCLUSIONS AI-generated medical information can achieve high decision-level accuracy while still lacking documentary support in a substantial proportion of outputs. Measuring evidence support and hallucination reveals important limitations that are not captured by agreement metrics alone. Retrieval-based evidence verification provides a reproducible and transparent approach for evaluating the reliability of AI-generated medical information, with direct relevance to digital health practice, evidence-based medicine, and medical informatics.

CLIP and BiomedCLIP are examples of vision-language foundation models and offer strong cross-modal embeddings; however, they are not optimized for fine-grained medical retrieval tasks, such as retrieving clinically relevant radiology reports using chest X-ray (CXR) image queries. To address this shortcoming, we propose a multi-task learning framework to fine-tune BiomedCLIP and evaluate improvements to CXR image-text retrieval. Using BiomedCLIP as the backbone, we incorporate a lightweight MLP projector head trained with a multi-task composite loss function that includes: (1) a binary cross-entropy loss to distinguish normal from abnormal CXR studies, (2) a supervised contrastive loss to reinforce intra-class consistency, and (3) a CLIP loss to maintain cross-modal alignment. Experimental results demonstrate that the fine-tuned model achieves more balanced and clinically meaningful performance across both image-to-text and text-to-image retrieval tasks compared to the pretrained BiomedCLIP and general-purpose CLIP models. Furthermore, t-SNE visualizations reveal clearer semantic clustering of normal and abnormal cases, demonstrating the model's enhanced diagnostic sensitivity. These findings highlight the value of domain-adaptive, multi-task learning for advancing cross-modal retrieval in biomedical applications.

Oral cancer has one of the lowest five-year survival rates among major cancer types. Therefore, early detection is crucial for histopathological confirmation. State-of-the-art methods reported in the literature largely analyze images only for oral cancer prediction. The use of deep learning networks for related tabular medical data remains unexplored for oral cancer and understudied in general. As part of our multimodal AI/ML approach toward reliable prediction of candidate lesions to biopsy, we describe our work in deep learning approaches on a fielded clinical structured text data in spreadsheet format (tabular data) on a subset comprising 1791 patients drawn from a large ongoing oral cancer study to classify patients with a cancerous lesion from those with a precancerous lesion (i.e., direct precursor to cancer). We compare two tabular deep learning methods and one conventional algorithm for the predictive data analysis. The experimental results on a hold-out test set demonstrate a promising performance for all models (Youden index > 0.6 and AUC > 0.9). In addition, we examine and analyze the interpretability of models. All models indicate that lesion characteristics are crucial predictive features. The insights and results obtained from this work would be valuable to the research community in application of AI/ML to biomedicine.

Biomedical data are inherently multimodal, capturing complementary aspects of a patient condition. Deep learning (DL) algorithms that integrate multiple biomedical modalities can significantly improve clinical decisionmaking, especially in domains where collecting data is not simple and data are highly heterogeneous. However, developing effective and reliable multimodal DL methods remains challenging, requiring large training datasets with paired samples from modalities of interest. An increasing number of de-identifed biomedical datasets are publicly accessible, though they still tend to be unimodal. For example, several publicly available skin lesion datasets aid automated dermatology clinical decision-making. Still, they lack annotated reports paired with the images, thereby limiting the advance and use of multimodal DL algorithms. This work presents a strategy exploiting real and synthesized data in a multimodal architecture that encodes finegrained text representations within image embeddings to create a robust representation of skin lesion data. Large language models (LLMs) are used to synthesize textual descriptions from image metadata that are subsequently paired with the original skin lesion images and used for model development. The architecture is evaluated on the classification of skin lesion images, considering nine internal and external data sources. The proposed multimodal representation outperforms the unimodal one on the classification of skin lesion images, achieving superior performance in every tested dataset.

Deep learning (DL) has transformed medical image classification; however, its efficacy is often limited by significant data imbalance due to far fewer cases (minority class) compared to controls (majority class). It has been shown that synthetic image augmentation techniques can simulate clinical variability, leading to enhanced model performance. We hypothesize that they could also mitigate the challenge of data imbalance, thereby addressing overfitting to the majority class and enhancing generalization. Recently, latent diffusion models (LDMs) have shown promise in synthesizing high-quality medical images. This study evaluates the effectiveness of a text-guided image-to-image LDM in synthesizing disease-positive chest X-rays (CXRs) and augmenting a pediatric CXR dataset to improve classification performance. We first establish baseline performance by fine-tuning an ImageNet-pretrained Inception-V3 model on class-imbalanced data for two tasks—normal vs. pneumonia and normal vs. bronchopneumonia. Next, we fine-tune individual text-guided image-to-image LDMs to generate CXRs showing signs of pneumonia and bronchopneumonia. The Inception-V3 model is retrained on an updated data set that includes these synthesized images as part of augmented training and validation sets. Classification performance is compared using balanced accuracy, sensitivity, specificity, F-score, Matthews correlation coefficient (MCC), Kappa, and Youden’s index against the baseline performance. Results show that the augmentation significantly improves Youden’s index (p<0.05) and markedly enhances other metrics, indicating that data augmentation using LDM-synthesized images is an effective strategy for addressing class imbalance in medical image classification.