Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:

Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_117_библиотеки_им_акад_М_И_Перельмана

.pdf
Скачиваний:
2
Добавлен:
15.09.2026
Размер:
15 Мб
Скачать
☆
Explainable AI: Methods, Frameworks, andTools forHealthcare 5.0
8. Abdulmalek S, Nasir A, Jabbar WA, Almuhaya MAM, Bairagi AK, Khan MAM et al (2022) IoT-based healthcare-monitoring system towards improving quality of life: a review. Healthcare (Basel) 10(10):1993. https://doi.org/10.3390/healthcare10101993
9. Pradhan B, Bhattacharyya S, Pal K (2021) IoT-based applications in healthcare devices. J Healthc Eng 2021:6632599. https://doi.org/10.1155/2021/6632599
10. Stockdale JE, Liu P, Colijn C (2022) The potential of genomics for infectious disease forecast­ing. Nat Microbiol 7(11):1736–1743. https://doi.org/10.1038/s41564- 022- 01233- 6
11. Garcelon N, Burgun A, Salomon R, Neuraz A (2020) Electronic health records for the diag­nosis of rare diseases. Kidney Int 97(4):676–686. https://doi.org/10.1016/j.kint.2019.11.037
12. Muralitharan S, Nelson W, Di S, McGillion M, Devereaux PJ, Barr NG etal (2021) Machine learning-based early warning systems for clinical deterioration: systematic scoping review. J Med Internet Res 23(2):e25187. https://doi.org/10.2196/25187
13. Lauritsen SM, Kristensen M, Olsen MV, Larsen MS, Lauritsen KM, Jørgensen MJ etal (2020) Explainable articial intelligence model to predict acute critical illness from electronic health records. Nat Commun 11(1):3852. https://doi.org/10.1038/s41467- 020- 17431- x
14. Norori N, Hu Q, Aellen FM, Faraci FD, Tzovara A (2021) Addressing bias in big data and AI for health care: a call for open science. Patterns (N Y) 2(10):100347. https://doi.org/10.1016/j.
patter.2021.100347
15. Pagano TP, Loureiro RB, Lisboa FVN, Peixoto RM, Guimarães GAS, Cruz GOR etal (2023) Bias and unfairness in machine learning models: a systematic review on datasets, tools, fair­ness metrics, and identication and mitigation methods. Big Data Cogn Comput 7(1):15.
https://doi.org/10.3390/bdcc7010015
16. Srinivasu PN, Sandhya N, Jhaveri RH, Raut R (2022) From blackbox to explainable AI in healthcare: existing tools and case studies. Mob Inf Syst 2022:1–20. https://doi.
org/10.1155/2022/8167821
17. Xu Q, Xie W, Liao B, Hu C, Qin L, Yang Z et al (2023) Interpretability of clinical decision support systems based on articial intelligence from technological and medical perspective: a systematic review. J Healthc Eng 2023:9919269. https://doi.org/10.1155/2023/9919269
18. Wang L, Chen X, Zhang L, Li L, Huang Y, Sun Y etal (2023) Articial intelligence in clinical decision support systems for oncology. Int J Med Sci 20(1):79–86. https://doi.org/10.7150/
ijms.77205
19. Zhang Y, Weng Y, Lund J (2022) Applications of explainable articial intelligence in diagnosis and surgery. Diagnostics (Basel) 12(2):237. https://doi.org/10.3390/diagnostics12020237
20. Yang G, Ye Q, Xia J (2022) Unbox the black-box for the medical explainable AI via multi­modal and multi-centre data fusion: a mini-review, two showcases and beyond. Inform Fusion 77:29–52. https://doi.org/10.1016/j.inffus.2021.07.016
21. Moradi M, Samwald M (2022) Deep learning, natural language processing, and explain­able articial intelligence in the biomedical domain. arXiv [cs.AI]. http://arxiv.org/
abs/2202.12678
22. Kwong JCC, Khondker A, Tran C, Evans E, Cozma AI, Javidan A etal (2022) Explainable articial intelligence to predict the risk of side-specic extraprostatic extension in pre­prostatectomy patients. Can Urol Assoc J 16(6):213–221. https://doi.org/10.5489/cuaj.7473
23. Fuhrman JD, Gorre N, Hu Q, Li H, El Naqa I, Giger ML (2022) A review of explainable and interpretable AI with applications in COVID-19 imaging. Med Phys 49(1):1–14. https://doi.
org/10.1002/mp.15359
24. Speith T (2022) A review of taxonomies of explainable articial intelligence (XAI) methods. In: 2022 ACM conference on fairness, accountability, and transparency. ACM, NewYork
25. Raza A, Tran KP, Koehl L, Li S (2022) Designing ECG monitoring healthcare system with federated transfer learning and explainable AI.Knowl Based Syst 236:107763. https://doi.
org/10.1016/j.knosys.2021.107763
26. Fan Z, Gong P, Tang S, Lee CU, Zhang X, Song P etal (2022) Joint localization and classica­tion of breast tumors on ultrasound images using a novel auxiliary attention-based framework. arXiv [eess.IV]. http://arxiv.org/abs/2210.05762
85
86
27. Sutton RT, Zai Ane OR, Goebel R, Baumgart DC (2022) Articial intelligence enabled auto­mated diagnosis and grading of ulcerative colitis endoscopy images. Sci Rep 12(1):2748.
https://doi.org/10.1038/s41598- 022- 06726- 2
28. Alsinglawi B, Alshari O, Alorjani M, Mubin O, Alnajjar F, Novoa M etal (2022) An explain­able machine learning framework for lung cancer hospital length of stay prediction. Sci Rep 12(1):607. https://doi.org/10.1038/s41598- 021- 04608- 7
29. Du Y, Rafferty AR, McAuliffe FM, Wei L, Mooney C (2022) An explainable machine learning- based clinical decision support system for prediction of gestational diabetes mel­litus. Sci Rep 12(1):1170. https://doi.org/10.1038/s41598- 022- 05112- 2
30. Severn C, Suresh K, Görg C, Choi YS, Jain R, Ghosh D (2022) A pipeline for the implementa­tion and visualization of explainable machine learning for medical imaging using radiomics features. Sensors (Basel) 22(14):5205. https://doi.org/10.3390/s22145205
31. Chaddad A, Daniel P, Zhang M, Rathore S, Sargos P, Desrosiers C etal (2022) Deep radiomic signature with immune cell markers predicts the survival of glioma patients. Neurocomputing 469:366–375. https://doi.org/10.1016/j.neucom.2020.10.117
32. Akula AR, Wang K, Liu C, Saba-Sadiya S, Lu H, Todorovic S etal (2022) CX-ToM: coun­terfactual explanations with theory-of-mind for enhancing human trust in image recognition models. iScience 25(1):103581. https://doi.org/10.1016/j.isci.2021.103581
33. van der Velden BHM, Kuijf HJ, Gilhuijs KGA, Viergever MA (2022) Explainable arti­cial intelligence (XAI) in deep learning-based medical image analysis. Med Image Anal 79:102470. https://doi.org/10.1016/j.media.2022.102470
34. Chaddad A, Lu Q, Li J, Katib Y, Kateb R, Tanougast C etal (2022) Explainable, domain­adaptive, and federated articial intelligence in medicine. arXiv [cs.CV]. http://arxiv.org/
abs/2211.09317
35. Bhattacharya P, Obaidat MS, Savaliya D, Sanghavi S, Tanwar S, Sadaun B (2022) Metaverse assisted telesurgery in healthcare 5.0: an interplay of blockchain and explainable AI. In: 2022 International conference on computer, information and telecommunication systems (CITS). IEEE
36. Paul M, Maglaras L, Ferrag MA, Almomani I (2023) Digitization of healthcare sector: a study on privacy and security concerns. ICT Express 9(4):571–588. https://doi.org/10.1016/j.
icte.2023.02.007
37. Guo P, Wang P, Zhou J, Jiang S, Patel VM (2021) Multi-institutional collaborations for improv­ing deep learning-based magnetic resonance image reconstruction using federated learning.
https://doi.org/10.48550/ARXIV.2103.02148
S. Pulipeti et al.
Explainable AI inDisease Diagnosis
PunamBedi, AnjaliThukral, andShivaniDhiman
Abstract With the advancement of Articial Intelligence (AI) techniques, intelli-
gent healthcare applications, such as early disease diagnosis, treatment of rare dis­eases, diagnosis using images, and clinical decision support systems, have become more ever-demanding. Disease Diagnosis (DD) is the most crucial area which needs early detection. In countries having voluminous populations and comparatively few healthcare professionals, there is a dire need for AI-enabled systems using Machine Learning (ML) and Deep Learning (DL) models for the fast and timely detection of diseases. Though ML-/DL-based systems have shown excellent performance across various domains, DD is still in the nascent stages of adaptability among medical practitioners. The leading cause is their lack of trust in black-box approaches that involve complex computations in their hidden layers. Explainable Articial Intelligence (XAI), a relatively new eld of AI, explains or interprets the recom­mendations generated by ML/DL models. Interpretability, tractability and explain­ability of machine-generated recommendations are essential in healthcare, especially in DD.This chapter discusses XAI post-hoc algorithms, such as Local Interpretable Model Agnostic Explanation, Partial Dependence Plot, Shapley Additive Explanations, and Gradient Class Activation Map, that explain ML-/DL-generated recommendations for DD.These explanations may help medical practitioners visu­alise the features contributing to algorithm-generated decisions. Consequently, it reduces errors during diagnosis and makes the system more trustworthy. The chap­ter explains the implementation of XAI methods using case studies on COVID-19 and Cancer diagnosis. XAI has shown promising outcomes. Yet, there is much scope for further exploration and research. The paper, therefore, discusses various limitations of XAI and challenges.
P. Bedi · S. Dhiman Department of Computer Science, University of Delhi, New Delhi, Delhi, India e-mail: pbedi@cs.du.ac.in; shivani@cs.du.ac.in
A. Thukral (*) Keshav Mahavidyalaya, University of Delhi, New Delhi, Delhi, India e-mail: athukral@keshav.du.ac.in
Ltd. 2024 R. Aluvalu et al. (eds.), Explainable AI in Health Informatics, Computational Intelligence Methods and Applications,
https://doi.org/10.1007/978-981-97-3705-5_5
87© The Author(s), under exclusive license to Springer Nature Singapore Pte
88
P. Bedi et al.
Keywords Explainable AI · Machine learning · Deep learning · Disease diagnosis
· Post-hoc methods
1 Introduction
Responsible Articial Intelligence [1] ensures that the path of Articial Intelligence (AI) leads to the development of society without any unfair bias. Responsible AI guarantees unbiasedness, explainability, fairness, privacy, and security in AI-based applications. Privacy and security in AI have been the top priorities for decades, whereas explainability has recently started receiving attention. Explainability ensures unbiasedness and fairness to AI-generated decisions.
AI has surpassed humans in the manner of time and is well-matched in perfor­mance in many daily tasks. Therefore, it has gained the attention of researchers worldwide. AI uses Machine Learning (ML) and Deep Learning (DL) algorithms to showcase human intelligence. However, a signicant concern remains intact—the black-box nature of these algorithms. Most of these algorithms are treated as black boxes due to the complex and extensive computations. There might be cases where the results or outcomes produced by the ML/DL algorithms are susceptible to the end users. The absence of any explanation or interpretation of these outcomes leads to doubts and distrust in AI systems. XAI, a subeld of AI (Fig.1), can ll the gap by providing explanations for the algorithmically generated outcomes.
XAI has been tried in various domains. Cyber security is a eld that prevents and protects against attacks in the cyber world, such as systems, programmes, and net­works. AI has immense potential to predict attacks and nd vulnerabilities in
Fig. 1 Articial intelligence and its subelds
Explainable AI inDisease Diagnosis
89
systems [2]. XAI can help to mitigate cyberattacks by providing explanations behind predictions. Another application area of XAI is object detection in autono­mous vehicles. AI techniques can be used to detect objects on roads to navigate trafc and take appropriate action based on the detection. However, the reasons behind the action taken (the detected objects and their positioning) are not explained to the end users by AI techniques. XAI can assist in explaining these reasons and enhance trust in AI systems [3]. Other elds where XAI contributes are automated ML [4], network intrusion [5], intelligent tutoring [6], building materials and con­struction [7], and chemistry [8].
While XAI is signicant for many industries, it is especially crucial for health­care because human lives are at risk. Lakhani etal. [9] performed a tuberculosis (TB) diagnosis from chest X-ray images. It was deduced in the paper that Convolutional Neural Network (CNN) architectures GoogleNet [10] and AlexNet [11] give a high accuracy in TB diagnosis, comparable to radiologists. However, despite AI virtually achieving 100% accuracy in many healthcare jobs, medical practitioners are usually reluctant to completely trust AI recommendations due to black-box computations in outcome generation. The survey paper [12] shows dis­trust of the general public in AI-based medicinal applications, such as surgery and radiology. Because of the distrust of the general public, AI has not been included in the workow of hospitals even though AI can signicantly reduce medical costs and provide faster and more personalised healthcare treatments. The requirement of explanations becomes more important where accuracy is a must in decision- making. A physician/medical practitioner needs a support mechanism to trust a disease pre­diction model. A model with 99% accuracy can produce 10,000 wrong predictions in onemillion cases which is a massive number. Thus, XAI can act as a support mechanism tool that allows a physician to seek reasons for the predictions. XAI has many applications in the healthcare sector, including AI-assisted Drug Design [13], AI-integrated healthcare conditions predictions [14], and Clinical Decision Support Systems (CDSS) [15]. One of the most signicant research areas in healthcare where XAI prominently contribute is Disease Diagnosis [16, 17].
The chapter is organised into seven sections. Various XAI taxonomies from the literature are discussed in Sect. 2, followed by a simple and comprehensive pro­posed XAI taxonomy. Section 3 explains different post-hoc XAI methods. Section
4 describes ML-/DL-based XAI methods that can be applied in healthcare, particu-
larly in disease diagnosis. Section 5 demonstrates two case studies of post-hoc methods: COVID-19 disease diagnosis using chest X-ray images and cancer diag­nosis using breast nuclei characteristics. Section 6 discusses the limitations and challenges of XAI.Finally, Sect. 7 concludes the chapter.
90
P. Bedi et al.
2 Taxonomy inExplainable AI (XAI)
Literature witnesses many classications based on different criteria applied to XAI methods, such as explanations derived through a particular sample (local) or whole dataset (global); and the time of producing explanations (before, during and after the training). This section discusses various taxonomies from literature and a pro­posed comprehensive and all-inclusive XAI taxonomy.
2.1 XAI Methods Classications
Intrinsic vs Model-Agnostic [18] Intrinsic methods are explainable by design and
possess accessibility to a model’s internal workings, such as weights. These are model-specic, and their application is limited to one specic model (say, logistic regression). In contrast, model-agnostic (post-hoc) methods are model-independent and do not understand the internal mechanics of models. They apply different tech­niques on input and model predictions to generate explanations. There are limita­tions to both types of methods. Intrinsic methods are simple in design and trade the performance with explanations. However, model-agnostic methods are applied after model prediction, preserving good prediction abilities without losing transparency. Model-agnostic methods struggle with highly complex datasets.
Local vs Global [19] Local and Global are the two categories of XAI methods based on the scope of their explanations. Local explanation-based methods generate explanations by analysing predictions of an individual input instance, whereas global explanations are produced by examining the entire dataset. Local techniques indicate the importance of input attributes to the prediction of a single instance, whereas global methods demonstrate the relevance of each input attribute to the overall predictions. Complex models have numerous parameters, making generalis­ing the feature importance locally and globally challenging.
Pre-modelling, In-modelling and Post-modelling [20] Pre-modelling does not involve AI models and focuses more on examining and comprehending the dataset than predictions. Insights gained from data via pre-modelling techniques can improve AI models’ effectiveness, explicability, and robustness. However, in­modelling methods generate explanations within the model whereas post-modelling generates explanations after the model training. In-modelling and post-modelling are similar to intrinsic and model-agnostic, respectively.
Perturbation vs Gradient Attribution [21] The perturbation attribution technique assesses the signicance of an input attribute in prediction by comparing its impact on the output with its presence or absence. They are model-agnostic methods. The main issue with these methods is the combinatorial explosion. In contrast, the
Explainable AI inDisease Diagnosis
Gradient-based attribution method evaluates feature importance by taking partial derivatives of output with respect to the input. These derivatives can be multiplied with the input to understand the effect of input on the outcome. Gradient-based attributions are model-specic methods.
Overlaps in criteria-based XAI classication create confusion for the readers. Therefore, an all-inclusive, comprehensive, and simple taxonomy is built and pre­sented below that incorporates other important categories which are absent in the existing taxonomies.
91
2.2 All-Inclusive XAI Taxonomy
The proposed all-inclusive XAI taxonomy provides a simplied and comprehensive classication of XAI methods. It is illustrated in Fig.2. All XAI methods can be broadly classied as pre-, in-, and post-modelling at the rst level. Further, the XAI methods are classied into feature-based (attribute) and gradient-based approaches. Feature-based approaches focus on the characteristics or attributes of a dataset, the relevance of features to an output, and sorting/ranking attributes as per their inu­ence on the outcome. Feature-based XAI methods have also emerged as a technol­ogy to extract/identify unapparent features. Under feature perturbation, approximation and range variations are subcategories. Approximation methods use techniques to nd an approximation of the predicted outcomes, whereas range vari­ation uses the range of attribute values to nd appropriate explanations. Gradient­based XAI methods work on the gradients computed in neural networks. Attribute propagation, a gradient-based method, assigns a relevance score to each attribute based on the gradients. The gradient method works on the computed gradients and applies a function (say average) on them, and the weighted gradient considers class­wise (labels) weights of the gradients to show the relevant input features. These criterion-based classications are represented with hollow rectangles in the XAI taxonomy (Fig.2). The leaf nodes in the taxonomy are XAI methods. The colour and shape of the nodes specify the type of scope of explanations as coded in the gure’s legend.
The most widely used XAI post-hoc methods are model-independent and com­patible with existing ML/DL algorithms that have shown excellent performance in human intelligence tasks. The widely used XAI methods are discussed in the next section.
92
Classification
Method
Method
P. Bedi et al.
Criteriabased
Fig. 2 All-inclusive XAI methods classication
Local ScopeXAI
Global ScopeXAI
Explainable AI inDisease Diagnosis
93
3 Post-hoc Methods inXAI
XAI methods have been applied in various AI elds such as computer vision [22,
23], natural language processing [24], and decision support systems [15]. For exam-
ple, in [15] the CDSS provided recommendations for gestational diabetes mellitus in pregnant women. Table1 lists various XAI methods applied in different datasets for explanations, in literature. It additionally lists the representation of explanations and the scope of explanations.
Local Interpretable Model Agnostic Explanation (LIME), Shapley Additive Explanations (SHAP), Partial Dependence Plot (PDP), and Gradient Class Activation Map (Grad-CAM) are some of the popular post-hoc based XAI algo­rithms that explain ML-/DL-generated recommendations. LIME and SHAP are approximation-based XAI methods used in various computer vision applications. PDP is a range variation method whereas Grad-CAM is a gradient-based XAI method.
Locally Interpretable Model-Agnostic Explanations (LIME) [18] LIME is an XAI method that generates local explanations independent of the AI model used for predictions. That means it analyses each instance’s outcome individually. New data instances are generated by varying values of input attributes for a specic instance (say x). Attribute values are increased/decreased near x to get a neighbourhood.
Table 1 XAI methods applied on different datasets listed with their output form and scope of their explanations
Literature reference
Y.Abdelwahab etal. [25]
J.Peng etal. [26] Relational
S.Liu etal. [24] Text NLIZE Post-hoc Local Attention heat maps Z.U. Ahmed
etal. [17]
H.Panwar etal. [23]
Y.Du etal. [15] Relational
R.Fong etal. [22]
R.Ying etal. [27] Graph GNNExplainer Intrinsic Global Graphs Y.Y. Jo etal. [28] Signals Saliency method Intrinsic Global Sensitivity map
Dataset format XAI methods
Short text (Tweets)
data
Mixed attributes
Image Grad-CAM Post-hoc Local Class-discriminative
data
Image Extreme
LIME Post-hoc Local Graph with
LIME Post-hoc Local Feature maps SHAP Post-hoc Global SHAP-generated
PDP Post-hoc Global PDP plots
PD Post-hoc Global PD prole LD Post-hoc Global LD prole AL Post-hoc Global AL prole
Explanation by example
SHAP Post-hoc Global Bar plots and
perturbation
Intrinsic/ post-hoc
Post-hoc Local Natural language
Post-hoc Local Saliency maps
Local/ global
Form of explanations
percentage
graphs
localisation map
and table
natural language
94
ε τλ
xLsfgg
x
()=()
+
()
,,
L
zz
()=() ()−()
()
′
∑
2
zx
xx
zMz
M
i
−−
()
 
′′
′
⊆
P. Bedi et al.
LIME uses an interpretable (surrogate) model (g) to get an approximation of the complex AI model. The possible interpretable models can be linear regression, deci­sion trees and their variation.
Equation (1) shows the optimisation function ε(x) of LIME. The rst term Ls(f, g, τx) looks for an approximation of the complex model f by the interpretable model g in the local neighbourhood τx. λ(g) is the complexity regularisation term and ensures the interpretable model stays simple to generate simple and clear explanations.
(1)
Let g be the linear regression, then, the loss function (Ls) can be calculated using Eq. (2), where f(z) is the class label and z′ is a new input instance.
sfgzfz gz
,,
ττ
x
x
′
(2)
LIME can be applied with varied data types, including tabular, textual, and image. However, LIME is a locally reliable XAI method as it creates a new dataset locally around a particular instance for explanation and cannot be generalised for global explanations.
Shapley Additive Explanations (SHAP) [29] It is based on Shapley values from cooperative game theory. These values are the measures of an individual or com­bined feature contribution to the outcome. The mathematical formula for SHAP is presented in Eq. (3).
fx
,
=
()
i
∑
!!
1
′
fz f
()
!
′
z
−
(3)
φi are the Shapley value of the input attribute (i), f is the black-box model, M is the total number of attributes in the dataset, z′ is the subset of attributes of which impor­tance is to be calculated and x is the input instance(s). SHAP iterates over all pos­sible subsets to capture the interaction between attributes. In the case of complex input such as images, each pixel is not treated as a feature; instead, a mapping is generally used to summarise them.
The model predictions are compared with and without feature(s) i to relate the importance of the feature(s) to the prediction. However, this computation is expen­sive as different combinations and permutations of input attributes are computed to identify their importance. Instead of calculating all combinations, the approximate Shapley values are used [30] to resolve this issue. SHAP uses various approxima­tion methods such as Kernel SHAP, Tree SHAP, and Deep SHAP.These approxima­tion methods use other existing models internally. For example, Kernel SHAP uses linear regression to nd approximate Shapley values [30].