Pub Date : 2025-06-04DOI: 10.1109/TAI.2025.3576087
Yang Yang;Chao Wang;Lei Gong;Min Wu;Zhenghua Chen;Xuehai Zhou
Knowledge distillation has become increasingly popular for training compact neural network models that can achieve comparable performance to larger models. In order to improve the effectiveness of knowledge distillation, enhancing the quality of the teacher knowledge is a crucial aspect to consider. While existing efforts have predominantly focused on optimizing the structure of teacher models and refining training procedures, we argue that there is untapped potential in further enhancing knowledge distillation through the augmentation of the teacher knowledge itself. In this article, we introduce FG-KD, a novel forward gradient-based framework specifically designed for augmenting teacher knowledge in knowledge distillation. FG-KD comprises two fundamental components: a feature reconstructor and a relation-aware enhancer. Both components employ a forward gradient-based approach to unlock the latent potential for enhancing teachers’ knowledge, thereby providing an enriched foundation for knowledge distillation. The feature reconstructor operates at the feature level, enabling the optimization of the teacher knowledge by enhancing the encoding of high-dimensional spaces. On the other hand, the relation-aware enhancer operates at the logit level, with a focus on identifying and reinforcing the interclass and intraclass relationships within the teacher knowledge. Through extensive experiments conducted on image recognition tasks, we demonstrate the effectiveness of FG-KD in improving the performance of various knowledge distillation techniques, regardless of the specific teacher–student model combinations.
{"title":"FG-KD: A Novel Forward Gradient-Based Framework for Teacher Knowledge Augmentation","authors":"Yang Yang;Chao Wang;Lei Gong;Min Wu;Zhenghua Chen;Xuehai Zhou","doi":"10.1109/TAI.2025.3576087","DOIUrl":"https://doi.org/10.1109/TAI.2025.3576087","url":null,"abstract":"Knowledge distillation has become increasingly popular for training compact neural network models that can achieve comparable performance to larger models. In order to improve the effectiveness of knowledge distillation, enhancing the quality of the teacher knowledge is a crucial aspect to consider. While existing efforts have predominantly focused on optimizing the structure of teacher models and refining training procedures, we argue that there is untapped potential in further enhancing knowledge distillation through the augmentation of the teacher knowledge itself. In this article, we introduce FG-KD, a novel forward gradient-based framework specifically designed for augmenting teacher knowledge in knowledge distillation. FG-KD comprises two fundamental components: a feature reconstructor and a relation-aware enhancer. Both components employ a forward gradient-based approach to unlock the latent potential for enhancing teachers’ knowledge, thereby providing an enriched foundation for knowledge distillation. The feature reconstructor operates at the feature level, enabling the optimization of the teacher knowledge by enhancing the encoding of high-dimensional spaces. On the other hand, the relation-aware enhancer operates at the logit level, with a focus on identifying and reinforcing the interclass and intraclass relationships within the teacher knowledge. Through extensive experiments conducted on image recognition tasks, we demonstrate the effectiveness of FG-KD in improving the performance of various knowledge distillation techniques, regardless of the specific teacher–student model combinations.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 1","pages":"439-454"},"PeriodicalIF":0.0,"publicationDate":"2025-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145898201","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-06-04DOI: 10.1109/TAI.2025.3575554
An-An Liu;Yadong Zhao;Xin Wen;Rihao Chang;Weizhi Nie
Recommender systems typically exhibit severe popularity bias, with a few highly popular items receiving excessive exposure. Most existing studies tackle this bias in static settings. However, they neglect the dynamic nature of real-world recommendation scenarios and lack a thorough analysis into the root causes of bias, which makes it challenging to accurately model and mitigate the dynamically changing popularity bias and capture genuine user preferences. To this end, we propose a causal disentanglement sequential recommendation model (CDSRec) based on time series analysis and hidden variable separation. Our model leverages Markov chains to analyze historical interaction data within sequential recommendations, capturing the dynamic variations of item popularity and user preferences. Employing causal inference, we disentangle the potential factors implicated in popularity bias. Specifically, user–item interactions are primarily driven by personalized demands and item popularity. Through empirical analysis from a temporal perspective, we reveal that popularity has both positive and negative impacts, and attribute them to stable intrinsic quality factors and dynamic external interference factors. We construct a causal directed acyclic graph to elucidate the temporal correlations among different factors. Subsequently, we utilize historical interaction sequences and item-related attributes as auxiliary information to explicitly disentangle these factors as hidden variables. By reformulating the objective function to optimize the sequential VAE framework, our model effectively mitigates the negative impact of external interference factors. Extensive experimental results on three real-world datasets demonstrate the superiority of our proposed model.
{"title":"Causal Disentanglement for Tackling Popularity Bias in Sequential Recommendation","authors":"An-An Liu;Yadong Zhao;Xin Wen;Rihao Chang;Weizhi Nie","doi":"10.1109/TAI.2025.3575554","DOIUrl":"https://doi.org/10.1109/TAI.2025.3575554","url":null,"abstract":"Recommender systems typically exhibit severe popularity bias, with a few highly popular items receiving excessive exposure. Most existing studies tackle this bias in static settings. However, they neglect the dynamic nature of real-world recommendation scenarios and lack a thorough analysis into the root causes of bias, which makes it challenging to accurately model and mitigate the dynamically changing popularity bias and capture genuine user preferences. To this end, we propose a causal disentanglement sequential recommendation model (CDSRec) based on time series analysis and hidden variable separation. Our model leverages Markov chains to analyze historical interaction data within sequential recommendations, capturing the dynamic variations of item popularity and user preferences. Employing causal inference, we disentangle the potential factors implicated in popularity bias. Specifically, user–item interactions are primarily driven by personalized demands and item popularity. Through empirical analysis from a temporal perspective, we reveal that popularity has both positive and negative impacts, and attribute them to stable intrinsic quality factors and dynamic external interference factors. We construct a causal directed acyclic graph to elucidate the temporal correlations among different factors. Subsequently, we utilize historical interaction sequences and item-related attributes as auxiliary information to explicitly disentangle these factors as hidden variables. By reformulating the objective function to optimize the sequential VAE framework, our model effectively mitigates the negative impact of external interference factors. Extensive experimental results on three real-world datasets demonstrate the superiority of our proposed model.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 1","pages":"426-438"},"PeriodicalIF":0.0,"publicationDate":"2025-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145898225","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-06-04DOI: 10.1109/TAI.2025.3576201
Xinhui Yu;Arvin Tashakori;Liang Zou;Z. Jane Wang
Most existing federated learning (FL) frameworks use deterministic models as the task model, which may suffer from overfitting due to small-scale data at client sides. Since Bayesian learning (BL) can quantify the uncertainty associated with both model parameters and prediction outcomes, there have been efforts to integrate BL with FL and the global objective is transformed into posterior approximation using Bayesian optimization. Variational inference is commonly used in such efforts which utilize the global distribution as the prior for the optimization of local Bayesian neural networks (BNNs) and thus eliminates the need for assigning specific prior distributions for clients. However, due to statistical heterogeneity across clients, the global distribution, representing the collective knowledge of all clients, may not be precise as client prior. To address this concern, we propose a federated Bayesian learning framework with personalized priors (pFedBL) where each client is assigned with a local BNN. Specifically, we first introduce a KL-divergence-based distribution aggregation scheme to ensure the effectiveness of the global distribution. Meanwhile, under the mild assumption that the server has access to a general unlabeled dataset, the server uses predictions as well as predictive uncertainty of these data, derived from local BNNs, to construct feature distributions. These distributions are then provided to clients for fine-tuning the global distribution, resulting in personalized priors. In addition, to ensure optimal integration of local and global data insights, we design an adaptive $zeta$ strategy in the local objective function to balance the log-likelihood estimation term and the KL divergence term. We provide theoretical analysis regarding the upper bound of the averaged generalization error for the proposed pFedBL and experimental results demonstrate its effectiveness on three datasets under different problem settings.
{"title":"pFedBL: Federated Bayesian Learning With Personalized Prior","authors":"Xinhui Yu;Arvin Tashakori;Liang Zou;Z. Jane Wang","doi":"10.1109/TAI.2025.3576201","DOIUrl":"https://doi.org/10.1109/TAI.2025.3576201","url":null,"abstract":"Most existing federated learning (FL) frameworks use deterministic models as the task model, which may suffer from overfitting due to small-scale data at client sides. Since Bayesian learning (BL) can quantify the uncertainty associated with both model parameters and prediction outcomes, there have been efforts to integrate BL with FL and the global objective is transformed into posterior approximation using Bayesian optimization. Variational inference is commonly used in such efforts which utilize the global distribution as the prior for the optimization of local Bayesian neural networks (BNNs) and thus eliminates the need for assigning specific prior distributions for clients. However, due to statistical heterogeneity across clients, the global distribution, representing the collective knowledge of all clients, may not be precise as client prior. To address this concern, we propose a federated Bayesian learning framework with personalized priors (pFedBL) where each client is assigned with a local BNN. Specifically, we first introduce a KL-divergence-based distribution aggregation scheme to ensure the effectiveness of the global distribution. Meanwhile, under the mild assumption that the server has access to a general unlabeled dataset, the server uses predictions as well as predictive uncertainty of these data, derived from local BNNs, to construct feature distributions. These distributions are then provided to clients for fine-tuning the global distribution, resulting in personalized priors. In addition, to ensure optimal integration of local and global data insights, we design an adaptive <inline-formula><tex-math>$zeta$</tex-math></inline-formula> strategy in the local objective function to balance the log-likelihood estimation term and the KL divergence term. We provide theoretical analysis regarding the upper bound of the averaged generalization error for the proposed pFedBL and experimental results demonstrate its effectiveness on three datasets under different problem settings.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 1","pages":"455-470"},"PeriodicalIF":0.0,"publicationDate":"2025-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145898256","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-06-04DOI: 10.1109/TAI.2025.3576336
Yang Wang;Ya-Hui Jia;Wei-Neng Chen;Yi Mei
Neural combinatorial optimization (NCO) has achieved remarkable performance in solving individual vehicle routing problems (VRPs) by leveraging attention mechanisms. However, when generalizing across different problems, these methods perform poorly because the hard parameter sharing models they adopted are unable to capture the commonalities and peculiarities of different problems. To address this limitation, we propose a novel multitask NCO method called the soft parameter sharing model (SPSM) that incorporates multiple independent attention modules and a gating network. SPSM allows the model to learn both universal patterns and individualized requirements without explicitly designating any module as shared or task-specific. When solving a specific VRP, the gating network may decide the importance of the characteristics learned by each attention module. Additionally, we adopt the maximum entropy reinforcement learning to maintain the diversity of the model in the training process, which can prevent the model from being greedy for some dominant tasks or only for the training tasks. Experimental results demonstrate that SPSM significantly enhances zero-shot generalization performance across ten unseen VRP variants and real-world benchmark instances.
{"title":"Soft Parameter Sharing Model for Cross-Problem Generalization in Vehicle Routing Problems","authors":"Yang Wang;Ya-Hui Jia;Wei-Neng Chen;Yi Mei","doi":"10.1109/TAI.2025.3576336","DOIUrl":"https://doi.org/10.1109/TAI.2025.3576336","url":null,"abstract":"Neural combinatorial optimization (NCO) has achieved remarkable performance in solving individual vehicle routing problems (VRPs) by leveraging attention mechanisms. However, when generalizing across different problems, these methods perform poorly because the hard parameter sharing models they adopted are unable to capture the commonalities and peculiarities of different problems. To address this limitation, we propose a novel multitask NCO method called the soft parameter sharing model (SPSM) that incorporates multiple independent attention modules and a gating network. SPSM allows the model to learn both universal patterns and individualized requirements without explicitly designating any module as shared or task-specific. When solving a specific VRP, the gating network may decide the importance of the characteristics learned by each attention module. Additionally, we adopt the maximum entropy reinforcement learning to maintain the diversity of the model in the training process, which can prevent the model from being greedy for some dominant tasks or only for the training tasks. Experimental results demonstrate that SPSM significantly enhances zero-shot generalization performance across ten unseen VRP variants and real-world benchmark instances.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 1","pages":"471-485"},"PeriodicalIF":0.0,"publicationDate":"2025-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145898261","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-06-03DOI: 10.1109/TAI.2025.3575548
Zuobin Xiong;Wei Li;Yingshu Li;Zhipeng Cai
Recent advancement in generative artificial intelligence (AI) influenced a broad area with successful applications across multiple domains, including computer vision, natural language processing, and the Internet of Things (IoT). However, many existing implementations rely on centralized architectures, which introduce security and privacy concerns while also increasing communication overhead. Limited research has explored the development of distributed generative models, particularly in scenarios where training data originate from various heterogeneous sources. To fill the gap, this article introduces a distributed generative model framework aimed at enhancing data generation in hierarchical IoT systems. The proposed framework supports distributed data generation across three distinct scenarios: feature-related data, label-related data, and feature-label nonrelated data. Furthermore, both synchronous and asynchronous update mechanisms are incorporated to accommodate diverse application requirements within IoT environments. Comprehensive experiments using simulated, image, and tabular datasets are conducted to assess the performance of the proposed framework in comparison with state-of-the-art methods. The results indicate that the framework effectively produces high-quality synthetic data while preserving the integrity of downstream tasks. Beyond large language models (LLMs), these findings suggest that generative AI has the potential to transform data generation in distributed IoT systems and be extended to a broader range of applications.
{"title":"Distributed Generative Model: A Data Synthesizing Framework for Multisource Heterogeneous Data","authors":"Zuobin Xiong;Wei Li;Yingshu Li;Zhipeng Cai","doi":"10.1109/TAI.2025.3575548","DOIUrl":"https://doi.org/10.1109/TAI.2025.3575548","url":null,"abstract":"Recent advancement in generative artificial intelligence (AI) influenced a broad area with successful applications across multiple domains, including computer vision, natural language processing, and the Internet of Things (IoT). However, many existing implementations rely on centralized architectures, which introduce security and privacy concerns while also increasing communication overhead. Limited research has explored the development of distributed generative models, particularly in scenarios where training data originate from various heterogeneous sources. To fill the gap, this article introduces a distributed generative model framework aimed at enhancing data generation in hierarchical IoT systems. The proposed framework supports distributed data generation across three distinct scenarios: feature-related data, label-related data, and feature-label nonrelated data. Furthermore, both synchronous and asynchronous update mechanisms are incorporated to accommodate diverse application requirements within IoT environments. Comprehensive experiments using simulated, image, and tabular datasets are conducted to assess the performance of the proposed framework in comparison with state-of-the-art methods. The results indicate that the framework effectively produces high-quality synthetic data while preserving the integrity of downstream tasks. Beyond large language models (LLMs), these findings suggest that generative AI has the potential to transform data generation in distributed IoT systems and be extended to a broader range of applications.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 1","pages":"399-411"},"PeriodicalIF":0.0,"publicationDate":"2025-06-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145898259","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-06-03DOI: 10.1109/TAI.2025.3575745
Gašper Beguš;Maksymilian Dąbkowski;Ryan Rhodes
The performance of large language models (LLMs) has recently improved to the point where models can perform well on many language tasks. We show here that—for the first time—the models can also generate valid metalinguistic analyses of language data. We outline a research program where the behavioral interpretability of LLMs on these tasks is tested via prompting. LLMs are trained primarily on text—as such, evaluating their metalinguistic abilities improves our understanding of their general capabilities and sheds new light on theoretical models in linguistics. We show that OpenAI’s [56] o1 vastly outperforms other models on tasks involving drawing syntactic trees and phonological generalization. We speculate that OpenAI o1’s unique advantage over other models may result from the model’s chain-of-thought mechanism, which mimics the structure of human reasoning used in complex cognitive tasks, such as linguistic analysis.
{"title":"Large Linguistic Models: Investigating LLMs’ Metalinguistic Abilities","authors":"Gašper Beguš;Maksymilian Dąbkowski;Ryan Rhodes","doi":"10.1109/TAI.2025.3575745","DOIUrl":"https://doi.org/10.1109/TAI.2025.3575745","url":null,"abstract":"The performance of large language models (LLMs) has recently improved to the point where models can perform well on many language tasks. We show here that—for the first time—the models can also generate valid metalinguistic analyses of language data. We outline a research program where the <italic>behavioral interpretability</i> of LLMs on these tasks is tested via prompting. LLMs are trained primarily on text—as such, evaluating their metalinguistic abilities improves our understanding of their general capabilities and sheds new light on theoretical models in linguistics. We show that OpenAI’s <xref>[56]</xref> o1 vastly outperforms other models on tasks involving drawing syntactic trees and phonological generalization. We speculate that OpenAI o1’s unique advantage over other models may result from the model’s <italic>chain-of-thought</i> mechanism, which mimics the structure of human reasoning used in complex cognitive tasks, such as linguistic analysis.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"6 12","pages":"3453-3467"},"PeriodicalIF":0.0,"publicationDate":"2025-06-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11022724","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145674827","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In knowledge-based visual question answering (KB-VQA), the answer can be naturally represented by translating visual object embedding referred by the question according to the cross-modality relation embedding related to both the question and the image. Though the triplet representation of cross-modality knowledge is plausible and proven effective, these methods often encounter two challenges: 1) The semantic gap between the image and the question makes it difficult to accurately embed the cross-modality relation; and 2) the visual objects in the question often have ambiguous references in the input image. To solve the above challenges, we propose the image-caption-question translating embeddings (ICQ-TransE), which more effectively models both the cross-modality relation and the head entity of visual objects. Specifically, for cross-modality relation embedding, the designed image-caption-question information transmission mechanism transmits the information flow from image to question through the caption bridge, where the caption simultaneously has the visual content and the textual form. With this powerful bridge, cross-modality information can be more effectively fused, resulting in more precisely encoded relation embeddings. For the visual object embedding, instead of using a fixed number of visual regions as the previous methods, the most relevant visual regions to the question are dynamically selected. Experimental results on OK-VQA and KRVQA challenging datasets verify the effectiveness of ICQ-TransE compared with multiple state-of-the-art methods for visual question answering with knowledge. Our code will be available at https://github.com/cmcv2022/ICQ-TransE.
{"title":"ICQ-TransE: LLM-Enhanced Image-Caption-Question Translating Embeddings for Knowledge-Based Visual Question Answering","authors":"Heng Liu;Boyue Wang;Xiaoyan Li;Yanfeng Sun;Yongli Hu;Baocai Yin","doi":"10.1109/TAI.2025.3575553","DOIUrl":"https://doi.org/10.1109/TAI.2025.3575553","url":null,"abstract":"In knowledge-based visual question answering (KB-VQA), the answer can be naturally represented by translating visual object embedding referred by the question according to the cross-modality relation embedding related to both the question and the image. Though the triplet representation of cross-modality knowledge is plausible and proven effective, these methods often encounter two challenges: 1) The semantic gap between the image and the question makes it difficult to accurately embed the cross-modality relation; and 2) the visual objects in the question often have ambiguous references in the input image. To solve the above challenges, we propose the image-caption-question translating embeddings (ICQ-TransE), which more effectively models both the cross-modality relation and the head entity of visual objects. Specifically, for cross-modality relation embedding, the designed image-caption-question information transmission mechanism transmits the information flow from image to question through the caption bridge, where the caption simultaneously has the visual content and the textual form. With this powerful bridge, cross-modality information can be more effectively fused, resulting in more precisely encoded relation embeddings. For the visual object embedding, instead of using a fixed number of visual regions as the previous methods, the most relevant visual regions to the question are dynamically selected. Experimental results on OK-VQA and KRVQA challenging datasets verify the effectiveness of ICQ-TransE compared with multiple state-of-the-art methods for visual question answering with knowledge. Our code will be available at <uri>https://github.com/cmcv2022/ICQ-TransE</uri>.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 1","pages":"412-425"},"PeriodicalIF":0.0,"publicationDate":"2025-06-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145898253","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-05-08DOI: 10.1109/tai.2025.3567961
Jialu Pi, Juan Maria Farina, Chieh-Ju Chao, Chadi Ayoub, Reza Arsanjani, Imon Banerjee
Mitigating population drift is vital for developing robust AI models for clinical use. While current methodologies focus on reducing demographic bias in disease predictions, they overlook the significant impact of chronic comorbidities. Addressing these complexities is essential to enhance predictive accuracy and reliability across diverse patient demographics, ultimately improving healthcare outcomes. We propose a causal reasoning framework to address selection bias in opportunistic screening for 1-year composite MACE risk using chest X-ray images. Training in high-risk primarily Caucasian patients (43% MACE event), the model was evaluated in a lower-risk emergency department setting (12.8% MACE event) and a relatively lower-risk external Asian patient population (23.81% MACE event) to assess selection bias effects. We benchmarked our approach against a high-performance disease classification model, a propensity score matching strategy, and a debiasing model for unknown biases. The causal+confounder framework achieved an AUC of 0.75 and 0.7 on Shift data and Shift external, outperforming baselines, and a comparable AUC of 0.7 on internal data despite penalties for confounders. It minimized disparities in confounding factors and surpassed traditional and state-of-the-art debiasing methods. Experimental data show that integrating causal reasoning and confounder adjustments in AI models enhances their effectiveness. This approach shows promise for creating fair and robust clinical decision support systems that account for population shifts, ultimately improving the reliability and ethical integrity of AI-driven clinical decision-making.
{"title":"Mitigating Bias in Opportunistic Screening for MACE with Causal Reasoning.","authors":"Jialu Pi, Juan Maria Farina, Chieh-Ju Chao, Chadi Ayoub, Reza Arsanjani, Imon Banerjee","doi":"10.1109/tai.2025.3567961","DOIUrl":"10.1109/tai.2025.3567961","url":null,"abstract":"<p><p>Mitigating population drift is vital for developing robust AI models for clinical use. While current methodologies focus on reducing demographic bias in disease predictions, they overlook the significant impact of chronic comorbidities. Addressing these complexities is essential to enhance predictive accuracy and reliability across diverse patient demographics, ultimately improving healthcare outcomes. We propose a causal reasoning framework to address selection bias in opportunistic screening for 1-year composite MACE risk using chest X-ray images. Training in high-risk primarily Caucasian patients (43% MACE event), the model was evaluated in a lower-risk emergency department setting (12.8% MACE event) and a relatively lower-risk external Asian patient population (23.81% MACE event) to assess selection bias effects. We benchmarked our approach against a high-performance disease classification model, a propensity score matching strategy, and a debiasing model for unknown biases. The causal+confounder framework achieved an AUC of 0.75 and 0.7 on Shift data and Shift external, outperforming baselines, and a comparable AUC of 0.7 on internal data despite penalties for confounders. It minimized disparities in confounding factors and surpassed traditional and state-of-the-art debiasing methods. Experimental data show that integrating causal reasoning and confounder adjustments in AI models enhances their effectiveness. This approach shows promise for creating fair and robust clinical decision support systems that account for population shifts, ultimately improving the reliability and ethical integrity of AI-driven clinical decision-making.</p>","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":" ","pages":""},"PeriodicalIF":0.0,"publicationDate":"2025-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12768338/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145914250","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：