Mitigating Algorithmic Bias in AI-Driven Cardiovascular Imaging for Fairer Diagnostics
<b>Background/Objectives</b>: The research addresses algorithmic bias in deep learning models for cardiovascular risk prediction, focusing on fairness across demographic and socioeconomic groups to mitigate health disparities. It integrates fairness-aware algorithms, susceptible carrier-...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2024-11-01
|
| Series: | Diagnostics |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2075-4418/14/23/2675 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1846124381765369856 |
|---|---|
| author | Md Abu Sufian Lujain Alsadder Wahiba Hamzi Sadia Zaman A. S. M. Sharifuzzaman Sagar Boumediene Hamzi |
| author_facet | Md Abu Sufian Lujain Alsadder Wahiba Hamzi Sadia Zaman A. S. M. Sharifuzzaman Sagar Boumediene Hamzi |
| author_sort | Md Abu Sufian |
| collection | DOAJ |
| description | <b>Background/Objectives</b>: The research addresses algorithmic bias in deep learning models for cardiovascular risk prediction, focusing on fairness across demographic and socioeconomic groups to mitigate health disparities. It integrates fairness-aware algorithms, susceptible carrier-infected-recovered (SCIR) models, and interpretability frameworks to combine fairness with actionable AI insights supported by robust segmentation and classification metrics. <b>Methods</b>: The research utilised quantitative 3D/4D heart magnetic resonance imaging and tabular datasets from the Cardiac Atlas Project’s (CAP) open challenges to explore AI-driven methodologies for mitigating algorithmic bias in cardiac imaging. The SCIR model, known for its robustness, was adapted with the Capuchin algorithm, adversarial debiasing, Fairlearn, and post-processing with equalised odds. The robustness of the SCIR model was further demonstrated in the fairness evaluation metrics, which included demographic parity, equal opportunity difference (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.037</mn></mrow></semantics></math></inline-formula>), equalised odds difference (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.026</mn></mrow></semantics></math></inline-formula>), disparate impact (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>1.081</mn></mrow></semantics></math></inline-formula>), and Theil Index (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.249</mn></mrow></semantics></math></inline-formula>). For interpretability, YOLOv5, Mask R-CNN, and ResNet18 were implemented with LIME and SHAP. Bias mitigation improved disparate impact (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.80</mn></mrow></semantics></math></inline-formula> to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.95</mn></mrow></semantics></math></inline-formula>), reduced equal opportunity difference (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.20</mn></mrow></semantics></math></inline-formula> to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.05</mn></mrow></semantics></math></inline-formula>), and decreased false favourable rates for males (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.0059</mn></mrow></semantics></math></inline-formula> to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.0033</mn></mrow></semantics></math></inline-formula>) and females (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.0096</mn></mrow></semantics></math></inline-formula> to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.0064</mn></mrow></semantics></math></inline-formula>) through balanced probability adjustment. <b>Results</b>: The SCIR model outperformed the SIR model (recovery rate: <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>1.38</mn></mrow></semantics></math></inline-formula> vs <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.83</mn></mrow></semantics></math></inline-formula>) with a <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mo>−</mo><mn>10</mn><mo>%</mo></mrow></semantics></math></inline-formula> transmission bias impact. Parameters (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>β</mi><mo>=</mo><mn>0.5</mn></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>δ</mi><mo>=</mo><mn>0.2</mn></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>γ</mi><mo>=</mo><mn>0.15</mn></mrow></semantics></math></inline-formula>) reduced susceptible counts to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>2.53</mn><mo>×</mo><msup><mn>10</mn><mrow><mo>−</mo><mn>12</mn></mrow></msup></mrow></semantics></math></inline-formula> and increased recovered counts to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>9.98</mn></mrow></semantics></math></inline-formula> by <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>t</mi><mo>=</mo><mn>50</mn></mrow></semantics></math></inline-formula>. YOLOv5 achieved high Intersection over Union (IoU) scores (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>94.8</mn><mo>%</mo></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>93.7</mn><mo>%</mo></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>80.6</mn><mo>%</mo></mrow></semantics></math></inline-formula> for normal, severe, and abnormal cases). Mask R-CNN showed <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>82.5</mn><mo>%</mo></mrow></semantics></math></inline-formula> peak confidence, while ResNet demonstrated a <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>10.4</mn><mo>%</mo></mrow></semantics></math></inline-formula> accuracy drop under noise. Performance metrics (IoU: <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.91</mn></mrow></semantics></math></inline-formula>–<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.96</mn></mrow></semantics></math></inline-formula>, Dice: <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.941</mn></mrow></semantics></math></inline-formula>–<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.980</mn></mrow></semantics></math></inline-formula>, Kappa: <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.95</mn></mrow></semantics></math></inline-formula>) highlighted strong predictive accuracy and reliability. <b>Conclusions</b>: The findings validate the effectiveness of fairness-aware algorithms in addressing cardiovascular predictive model biases. The integration of fairness and explainable AI not only promotes equitable diagnostic precision but also significantly reduces diagnostic disparities across vulnerable populations. This reduction in disparities is a key outcome of the research, enhancing clinical trust in AI-driven systems. The promising results of this study pave the way for future work that will explore scalability in real-world clinical settings and address limitations such as computational complexity in large-scale data processing. |
| format | Article |
| id | doaj-art-a73526fea85c48ae922399637224f377 |
| institution | Kabale University |
| issn | 2075-4418 |
| language | English |
| publishDate | 2024-11-01 |
| publisher | MDPI AG |
| record_format | Article |
| series | Diagnostics |
| spelling | doaj-art-a73526fea85c48ae922399637224f3772024-12-13T16:24:37ZengMDPI AGDiagnostics2075-44182024-11-011423267510.3390/diagnostics14232675Mitigating Algorithmic Bias in AI-Driven Cardiovascular Imaging for Fairer DiagnosticsMd Abu Sufian0Lujain Alsadder1Wahiba Hamzi2Sadia Zaman3A. S. M. Sharifuzzaman Sagar4Boumediene Hamzi5IVR Low-Carbon Research Institute, Chang’an University, Xi’an 710018, ChinaInstitute of Health Sciences Education, Queen Mary University, London E1 4NS, UKLaboratoire de Biotechnologie Santé et Environnement, Department of Biology, University of Blida, Blida 09000, AlgeriaInstitute of Health Sciences Education, Queen Mary University, London E1 4NS, UKDepartment of AI and Robotics, Sejong University, Seoul 05006, Republic of KoreaDepartment of Computing and Mathematical Sciences, California Institute of Technology, Pasadena, CA 91125, USA<b>Background/Objectives</b>: The research addresses algorithmic bias in deep learning models for cardiovascular risk prediction, focusing on fairness across demographic and socioeconomic groups to mitigate health disparities. It integrates fairness-aware algorithms, susceptible carrier-infected-recovered (SCIR) models, and interpretability frameworks to combine fairness with actionable AI insights supported by robust segmentation and classification metrics. <b>Methods</b>: The research utilised quantitative 3D/4D heart magnetic resonance imaging and tabular datasets from the Cardiac Atlas Project’s (CAP) open challenges to explore AI-driven methodologies for mitigating algorithmic bias in cardiac imaging. The SCIR model, known for its robustness, was adapted with the Capuchin algorithm, adversarial debiasing, Fairlearn, and post-processing with equalised odds. The robustness of the SCIR model was further demonstrated in the fairness evaluation metrics, which included demographic parity, equal opportunity difference (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.037</mn></mrow></semantics></math></inline-formula>), equalised odds difference (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.026</mn></mrow></semantics></math></inline-formula>), disparate impact (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>1.081</mn></mrow></semantics></math></inline-formula>), and Theil Index (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.249</mn></mrow></semantics></math></inline-formula>). For interpretability, YOLOv5, Mask R-CNN, and ResNet18 were implemented with LIME and SHAP. Bias mitigation improved disparate impact (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.80</mn></mrow></semantics></math></inline-formula> to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.95</mn></mrow></semantics></math></inline-formula>), reduced equal opportunity difference (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.20</mn></mrow></semantics></math></inline-formula> to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.05</mn></mrow></semantics></math></inline-formula>), and decreased false favourable rates for males (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.0059</mn></mrow></semantics></math></inline-formula> to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.0033</mn></mrow></semantics></math></inline-formula>) and females (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.0096</mn></mrow></semantics></math></inline-formula> to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.0064</mn></mrow></semantics></math></inline-formula>) through balanced probability adjustment. <b>Results</b>: The SCIR model outperformed the SIR model (recovery rate: <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>1.38</mn></mrow></semantics></math></inline-formula> vs <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.83</mn></mrow></semantics></math></inline-formula>) with a <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mo>−</mo><mn>10</mn><mo>%</mo></mrow></semantics></math></inline-formula> transmission bias impact. Parameters (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>β</mi><mo>=</mo><mn>0.5</mn></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>δ</mi><mo>=</mo><mn>0.2</mn></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>γ</mi><mo>=</mo><mn>0.15</mn></mrow></semantics></math></inline-formula>) reduced susceptible counts to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>2.53</mn><mo>×</mo><msup><mn>10</mn><mrow><mo>−</mo><mn>12</mn></mrow></msup></mrow></semantics></math></inline-formula> and increased recovered counts to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>9.98</mn></mrow></semantics></math></inline-formula> by <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>t</mi><mo>=</mo><mn>50</mn></mrow></semantics></math></inline-formula>. YOLOv5 achieved high Intersection over Union (IoU) scores (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>94.8</mn><mo>%</mo></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>93.7</mn><mo>%</mo></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>80.6</mn><mo>%</mo></mrow></semantics></math></inline-formula> for normal, severe, and abnormal cases). Mask R-CNN showed <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>82.5</mn><mo>%</mo></mrow></semantics></math></inline-formula> peak confidence, while ResNet demonstrated a <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>10.4</mn><mo>%</mo></mrow></semantics></math></inline-formula> accuracy drop under noise. Performance metrics (IoU: <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.91</mn></mrow></semantics></math></inline-formula>–<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.96</mn></mrow></semantics></math></inline-formula>, Dice: <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.941</mn></mrow></semantics></math></inline-formula>–<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.980</mn></mrow></semantics></math></inline-formula>, Kappa: <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>0.95</mn></mrow></semantics></math></inline-formula>) highlighted strong predictive accuracy and reliability. <b>Conclusions</b>: The findings validate the effectiveness of fairness-aware algorithms in addressing cardiovascular predictive model biases. The integration of fairness and explainable AI not only promotes equitable diagnostic precision but also significantly reduces diagnostic disparities across vulnerable populations. This reduction in disparities is a key outcome of the research, enhancing clinical trust in AI-driven systems. The promising results of this study pave the way for future work that will explore scalability in real-world clinical settings and address limitations such as computational complexity in large-scale data processing.https://www.mdpi.com/2075-4418/14/23/2675cardiovascular risk predictionalgorithmic biasfairness-aware AIdemographic fairnessadversarial debiasingSCIR model |
| spellingShingle | Md Abu Sufian Lujain Alsadder Wahiba Hamzi Sadia Zaman A. S. M. Sharifuzzaman Sagar Boumediene Hamzi Mitigating Algorithmic Bias in AI-Driven Cardiovascular Imaging for Fairer Diagnostics Diagnostics cardiovascular risk prediction algorithmic bias fairness-aware AI demographic fairness adversarial debiasing SCIR model |
| title | Mitigating Algorithmic Bias in AI-Driven Cardiovascular Imaging for Fairer Diagnostics |
| title_full | Mitigating Algorithmic Bias in AI-Driven Cardiovascular Imaging for Fairer Diagnostics |
| title_fullStr | Mitigating Algorithmic Bias in AI-Driven Cardiovascular Imaging for Fairer Diagnostics |
| title_full_unstemmed | Mitigating Algorithmic Bias in AI-Driven Cardiovascular Imaging for Fairer Diagnostics |
| title_short | Mitigating Algorithmic Bias in AI-Driven Cardiovascular Imaging for Fairer Diagnostics |
| title_sort | mitigating algorithmic bias in ai driven cardiovascular imaging for fairer diagnostics |
| topic | cardiovascular risk prediction algorithmic bias fairness-aware AI demographic fairness adversarial debiasing SCIR model |
| url | https://www.mdpi.com/2075-4418/14/23/2675 |
| work_keys_str_mv | AT mdabusufian mitigatingalgorithmicbiasinaidrivencardiovascularimagingforfairerdiagnostics AT lujainalsadder mitigatingalgorithmicbiasinaidrivencardiovascularimagingforfairerdiagnostics AT wahibahamzi mitigatingalgorithmicbiasinaidrivencardiovascularimagingforfairerdiagnostics AT sadiazaman mitigatingalgorithmicbiasinaidrivencardiovascularimagingforfairerdiagnostics AT asmsharifuzzamansagar mitigatingalgorithmicbiasinaidrivencardiovascularimagingforfairerdiagnostics AT boumedienehamzi mitigatingalgorithmicbiasinaidrivencardiovascularimagingforfairerdiagnostics |