Explainable, trustworthy, and ethical machine learning for healthcare: A survey

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Few researchers have focused on the problem of evaluating explainable machine learning (XML).

“Despite the increasing research on the said topic, few researchers focused on the problem of evaluating XML.”

May combine two claims.

DL models have a black-box nature and lack an underlying theoretical foundation behind their decisions.

“However, these models have black-box nature and lack underlying theoretical foundation behind their decisions [8]”

• cites [8] Begoli, 2019, The need for uncertainty quantification in machine-assisted medical decision making, Nat. Mach. Intell. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

The human-machine combination is a powerful tool to facilitate patients and develop high-quality treatments.

“This human-machine combination is a powerful tool to facilitate the patients and develop high-quality treatments [21].”

• cites [21] M.A. Ahmad, C. Eckert, A. Teredesai, Interpretable machine learning in healthcare, in: Proceedings of the 2018 ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics, 2018, pp. 559–560. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Explainable techniques allow researchers or end-users to get involved in improving the performance of the algorithm and to trust the prediction results.

“The explainable techniques allow the researchers or end-users to get involved in improving the performance of the algorithm and to trust the prediction results [44].”

• cites [44] Caruana, 2015, Intelligible models for HealthCare: Predicting pneumonia risk and hospital 30-day readmissionNot checked yet

  Not checked in the free run.

May combine two claims.

Explanations of ML systems can help improve prescribed medicines, facilities provided to patients in hospitals, and health monitoring systems.

“These insights can help improve the prescribed medicines, facilities provided to patients in hospitals [46], and health monitoring systems [47].”

• cites [46] Guidotti, 2018, A survey of methods for explaining black box models, ACM Comput. Surv. sourceNot checked yet

  Not checked in the free run.

• cites [47] Doshi-Velez, 2017Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

Evaluation of explanations is a complicated task because explainability is non-monolithic and subjective.

“Due to the non-monolithic concept and subjective nature of explainability, evaluation of explanations is a complicated task.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Function-based evaluation does not require humans in the loop and works appropriately when human-based or application-based evaluations have already been performed.

“This approach does not require humans in the loop (layperson or domain expert).It works appropriately when human-based or application-based evaluations have already been performed.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Black-box methodology provides local explanations of a model for a prediction.

“The blackbox methodology provides local explanations of a model for a prediction [41].”

• cites [41] Chen, 2021, Ethical machine learning in healthcare, Annu. Rev. Biomed. Data Sci. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Human-based evaluation does not require a costly experimental environment and domain expert for testing.

“However, the difference is that it does not require a costly experimental environment and domain expert for testing.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

An XML technique usually explains in a human-understandable way how features of data relate to prediction results.

“An XML technique usually explains in a human-understandable way how the feature of data relates to prediction results, i.e., what features of X-ray images a model learns to detect the fractures.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Explainability of black-box models is required to attain fair and trustworthy healthcare decisions.

“It is now evident that the explainability of black-box models is required to attain fair and trustworthy healthcare decisions.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Low translucency has the advantage of more compact explanation results.

“Low translucency has the advantage of more compact explanation results.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Application-based evaluation helps in evaluating an explanation in real-time practical scenarios.

“This technique helps in evaluating the explanation in real-time practical scenarios.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The properties of good quality XML methods are mostly required for explanations of any black-box model.

“These properties are mostly required for explanations of any black-box model, however, we are describing these in terms of healthcare domain.”

• cites [51] Işın, 2016, Review of MRI-based brain tumor image segmentation using deep learning methods, Procedia Comput. Sci. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Robnik et al. listed properties of a good quality XML method.

“Robnik et al. [51] listed some properties of good quality XML method.”

• cites [51] Işın, 2016, Review of MRI-based brain tumor image segmentation using deep learning methods, Procedia Comput. Sci. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Model-based explanation methods are well developed and implemented for healthcare applications.

“The model-based explanation methods are well developed and implemented for healthcare applications (further discussed in Sec IV-D).”

• cites [50] Rasheed, 2020, Machine learning for predicting epileptic seizures using EEG signals: A review, IEEE Rev. Biomed. Eng.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Medical end-users likely do not have enough knowledge to understand complex mathematical explanations, so rule-based, textual, or visualization-based explanations are required for them.

“For end-users of the medical field, i.e., clinicians, radiologists, pathologists, neurologists, and patients, it is more likely that they do not have enough knowledge of understanding complex mathematical explanations.So, for them, rulebased, textual, or visualization-based explanations are required.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Human-based evaluation can be tested with laypersons and helps generalize findings because more testers are easily available.

“One can test the explanations with laypersons, and it helps to generalize the findings as the more number of testers (laypersons) are easily available.”

Probably not a substantial claim: may not pass S3 (load-bearing).

High translucency allows the explanation method to gather information from more internal parameters of a model.

“High translucency allows the explanation method to gather information from more internal parameters of a model.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Explanation provides a fair analysis of model architecture and learned parameters.

“Thus, the explanation provides a fair analysis of model architecture and learned parameters [42].”

• cites [42] Levin, 2021, Machine-learning-based hospital discharge predictions can support multidisciplinary rounds and decrease hospital length-of-stay, BMJ Innov. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Domain adaptable outputs concern how an explainable model represents its explanation according to the application domain and end-users.

“It is how an explainable model represents its explanation according to the application domain and end-users.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Translucency of XML represents how much the explaining method depends on the internal architecture of the ML model.

“It represents how much the explaining method depends on the internal architecture of the ML model, i.e., learnable parameters.”

• AssumptionRests on an assumption: There is no definite method to calculate the properties of good quality XML methods.

Probably not a substantial claim: may not pass S3 (load-bearing).

The signal method is used to highlight the patterns of data that activate the neurons of higher layers by back-propagating a signal from the last layer of the network to the input layer.

“The signal method is used to highlight the patterns of data that activate the neurons of higher layers.This can be done by back-propagating a signal from the last layer of the network to the input layer.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

For a generalized model for clinical use, low translucency is desirable; for a patient-specific application, the explanation should be more detailed and accurate, leading towards high translucency.

“To get a generalized model for clinical use low translucency is desirable and for a patient-specific application, the explanation should be more detailed and accurate that can lead towards high translucency.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Adoptability illustrates the variety of ML methods for which an explanation technique can be applied.

“It illustrates the variety of ML methods for which an explanation technique can be applied.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

A good explanation should be accurate especially in disease prediction; low accuracy is acceptable if the performance of the ML model is also low.

“A good explanation should be accurate specially in case of disease prediction.Low value of accuracy is acceptable if the performance of ML model is also low.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

LIME's local explanation allows the end-user to determine which feature is important for the precise prediction and how perturbation in a feature affects the prediction results.

“This local explanation allows the end-user to determine which feature is important for the precise prediction and how the perturbation in feature affects the prediction results.”

• cites [64] Samek, 2021, Explaining deep neural networks and beyond: A review of methods and applications, Proc. IEEE sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

Deep learning techniques are precise in terms of performance but their decisions are difficult to explain because of their complex model architecture.

“These techniques are precise in terms of performance but their decisions are difficult to explain because of their complex model architecture.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Saliency methods and feature attribution are the two broad categories of methods applied for explanation of deep learning models.

“Saliency methods and feature attribution (FA) are the two broad categories of methods applied for explanation of DL models.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Fidelity is a property of explanation that shows the precision of approximating the decision of an ML method.

“Fidelity is a property of explanation that shows the precision of approximating the decision of ML method.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Explanation methods of complex deep neural networks have high translucency and thus cannot be applicable for other models.

“Explanation methods of complex deep neural networks (DNN) have high translucency thus can not be applicable for other models.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Saliency methods produce the explanation by presenting important feature maps of each data sample.

“Saliency methods produce the explanation by presenting important feature maps of each data sample.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Samek et al. showed that heatmaps produced by sensitivity analysis are much more noisy compared to heatmaps generated using layer-wise relevance propagation.

“They showed that the heatmaps produced by SA are much more noisy compared to the heatmaps generated using LRP [65].”

• cites [65] Bohlin, 2006Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Gradient-based saliency methods reveal how the output of the model changes with a small change in the input and are computationally efficient because of a single pass of input through the network.

“Gradient-based saliency methods reveal how the output of the model changes with a small change in the input.These methods are computationally efficient because of a single pass of input (forward and backward) through the network.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

It is crucial for a typical ML/DL empowered healthcare system to be fully trusted by both clinicians and patients for getting the real impact of such systems.

“It is crucial for a typical ML/DL empowered healthcare system to be fully trusted by both the clinicians and patients for getting the real impact of such systems.”

Probably not a substantial claim: may not pass S3 (load-bearing).

LIME produces an explanation list that shows the contribution of individual features to the prediction.

“LIME produces an explanation list that shows the contribution of individual features to the prediction.”

• cites [64] Samek, 2021, Explaining deep neural networks and beyond: A review of methods and applications, Proc. IEEE sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The OECD and European Commission's AI High-Level Expert Group principles can be used for sustaining trust and trustworthiness in AI.

“The literature refers to two sets of popular principles that have been outlined by the Organisation for Economic Co-operation and Development (OECD) [67] and European Commission's AI High-Level Expert Group (HLEG) [68] that can be used for sustaining trust and trustworthiness in AI.”

• cites [67] Holzinger, 2022, Information fusion as an integrative cross-cutting enabler to achieve robust, explainable, and trustworthy medical artificial intelligence, Inf. Fusion sourceNot checked yet

  Not checked in the free run.

• cites [68] Mirsky, 2019, CT-GAN: Malicious tampering of 3D medical imagery using deep learningNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Techniques with low translucency are applicable to a wide range of ML methods.

“The techniques with low translucency are applicable to a wide range of ML methods.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The feature attribution method decomposes each value produced from each neuron of the output layer according to the contributions made by the individual dimensions of an input sample.

“The feature attribution method decomposes each value produced from each neuron of the output layer according to the contributions made by the individual dimensions of an input sample.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

A high concentration of cfDNA in plasma and CNV in chromosomes affected the pathogenesis of cancer cases.

“They concluded that a high concentration of cfDNA in plasma and CNV in chromosomes affected the pathogenesis of cancer cases.”

• cites [118] Bengio, 2010, Decision trees do not generalize to new variations, Comput. Intell. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Intraoperative urine output, IV fluid infusion, blood product transfusion, and dynamic changes of hemodynamic features are significant causes of cardiac surgery-associated acute kidney injury.

“Using SHAP values they described that the intraoperative urine output, IV fluid infusion, blood product transfusion, and dynamic changes of hemodynamic features are significant causes of injury.”

• cites [113] Palatnik de Sousa, 2019, Local interpretable model-agnostic explanations for classification of lymph node metastases, Sensors sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

CNN learned the patterns of peripheral lesion of DN glomeruli for DN detection.

“They analyzed the decision of CNN using LIME and described that CNN learned the patterns of peripheral lesion of DN glomeruli for DN detection.”

• cites [112] Jo, 2020, Deep learning detection of informative features in tau PET for Alzheimer’s disease classification, BMC BioinformaticsNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Adversarial robustness can be defined as the survivability of ML-based systems to adversarial attacks.

“The adversarial robustness can be defined as the survivability of ML-based systems to adversarial attacks.”

Probably not a substantial claim: may not pass S3 (load-bearing).

These factors were not revealed using traditional techniques.

“They also stated that these factors were not revealed using traditional techniques.”

• cites [113] Palatnik de Sousa, 2019, Local interpretable model-agnostic explanations for classification of lymph node metastases, Sensors sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

DLIME shows better results than LIME.

“They concluded that the DLIME shows better results than LIME.”

• cites [116] Pianpanit, 2019Not checked yet

  Not checked in the free run.

May combine two claims.

The majority of the guidelines in both the OECD and HLEG frameworks mainly focus on robustness, safety, security, explainability, and fairness and are therefore key requirements for building trustworthy AI systems.

“It is worth noting that the majority of the guidelines in both aforementioned frameworks mainly focus on the AI aspects of robustness, safety, security, explainability, and fairness and are therefore the key requirements for building trustworthy AI systems.”

• cites [67] Holzinger, 2022, Information fusion as an integrative cross-cutting enabler to achieve robust, explainable, and trustworthy medical artificial intelligence, Inf. Fusion sourceNot checked yet

  Not checked in the free run.

• cites [68] Mirsky, 2019, CT-GAN: Malicious tampering of 3D medical imagery using deep learningNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The robustness of ML/DL models is defined along three dimensions: robustness to security threats, robustness to distribution shifts, and data imperfections.

“The safe and robust ML is a broad term and we define the robustness of the ML/DL models along three dimensions, i.e., robustness to security threats, robustness to distribution shifts, and data imperfections.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Instability of generated explanations due to perturbation addition and random feature selection is a problem in medical computer-aided diagnosis systems.

“Zafar et al. [116] pointed out the problem of instability of generated explanations due to perturbation addition and random feature selection in the medical computer-aided diagnosis (CAD) systems.”

• cites [116] Pianpanit, 2019Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Trustworthy ML can only be possible by addressing challenges related to privacy, fairness, explainability, security, and robustness.

“From the figure, it is evident that the trustworthy ML can only be possible by addressing challenges related to privacy, fairness, explainability, security, and robustness.”

May combine two claims.

ML systems are not safe, secure, and robust.

“The literature suggests that ML systems are not safe, secure, and robust.”

• cites [32] Qayyum, 2021, Secure and robust machine learning for healthcare: A survey, IEEE Rev. Biomed. Eng. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Implementation of P2V improves predictive model working efficiency and increases explainability.

“The implementation of P2V improves the predictive model working efficiency and also increase the explainability of these models.”

• cites [122] H. Jung, Y. Oh, Towards better explanations of class activation mapping, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1336–1344. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

IG works on two axioms: sensitivity and implementation variance.

“This algorithm works on two axioms sensitivity and implementation variance.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

To resolve the problem of toxic drugs, a chemist needs the structural element causing the problem.

“To resolve the problem of toxic drugs a chemist need the structural element which is causing the problem.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

IG does not require any changes to the primordial deep neural network.

“IG does not require any changes to the primordial deep neural network.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

HSCNN out-performed CNN with a high prediction score and explainable results.

“They also compared the performance of their proposed model with CNN and concluded that the HSCNN out-performed the CNN with a high prediction score and explainable results.”

• cites [121] Guisan, 2002, Generalized linear and generalized additive models in studies of species distributions: setting the scene, Ecol. Model. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

IG can be implemented for any kind of model, i.e., image.

“IG can be implemented for any kind of model, i.e., image.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In drug development, classification of toxic and non-toxic drugs is not enough.

“In the drug development classification of toxic and non-toxic drugs is not enough.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

If the model outcome has changed notably with perturbations, it shows that the feature has a high contribution to the prediction.

“If the model outcome has changed notably with perturbations, it shows us that the feature has a high contribution to the prediction.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The network is sensitive to intensity and sphericity in coherence with domain information.

“They showed that the network is sensitive to intensity and sphericity in coherence with domain information.”

• cites [119] Mohnen, 2020, Healthcare expenditure prediction with neighbourhood variables–A random forest model, Stat. Politics Policy sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

HIPAA act assures the privacy of the medical data of patients.

“For example, the health insurance and portability and accountability (HIPAA) act assures the privacy of the medical data of patients.”

• cites [157] Panesar, 2019Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

ML algorithms require access to as much data as possible to improve the precision and accuracy of the outcome.

“However, ML algorithms requires access to as much data as possible to improve the precision and accuracy of the outcome.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

ML-based healthcare products are manufactured by Western companies and developed and tested on Caucasian data.

“Another problem is that ML-based healthcare products are manufactured by Western companies and these products are developed and tested on Caucasian data.”

Classified as probably not a substantial claim.

Explanation of data is required to check for these biases.

“Explanation of the data, as we proposed in the pipeline of explainable ML presented in Figure 3, in the first place is required to check for these biases.”

Probably not a substantial claim: may not pass S4 (based).

The quality of training data has a high impact on the performance of ML algorithms.

“Against this background, the quality of training data has a high impact on the performance of ML algorithms.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Imbalance class data is one of the common data-related problems in supervised training of ML/DL models.

“Imbalance class data is one of the common data-related problems that occur in the supervised training of ML/DL models.”

May combine two claims.

This bias in the availability of facilities is reflected in the data and can cause biased decisions of ML algorithms.

“This bias in the availability of facilities is also reflected in the data and can cause the biased decision of ML algorithms [155].”

• cites [155] Bohr, 2020Not checked yet

  Not checked in the free run.

May combine two claims.

More healthcare facilities are available for wealthy people, making it less likely that low-income people are able to access advanced technological treatments.

“It is a common practice that more healthcare facilities are available for wealthy people which makes it less likely that low-income people are able to access advanced technological treatments.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

AI/ML can improve care quality, cut overall costs, reduce or eliminate diagnostic errors, and improve the process of predicting disease.

“it can improve care quality, cut the overall costs, reduce or even eliminate diagnostic errors and improve the process of predicting disease.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

ML models learn the latent variable of data to deduce predictions.

“ML models learn the latent variable of data to deduce the predictions.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Protecting the privacy of patients and the confidentiality of their data is one of the fundamentals of ethical healthcare.

“Protecting the privacy of patients and the confidentiality of their data is one of the fundamentals of ethical healthcare.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

HIPAA's policy standards are designed to improve the healthcare systems and mandate all healthcare organizations to protect medical information.

“HIPAA's policy standards are designed to improve the healthcare systems and mandate it for all healthcare organizations to protect medical information [157].”

• cites [157] Panesar, 2019Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Integration of AI/ML into healthcare practice and clinical applications promises to provide substantial improvements to the healthcare sector.

“The integration of AI/ML into healthcare practice and clinical applications promises to provide substantial improvements to the healthcare sector.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Imbalance class data arises due to nonuniform distribution of samples among classes.

“This problem arises due to the nonuniform distribution of samples among classes.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Biases in outcomes of models used for healthcare services may have profound consequences.

“Biases in outcomes of models used for healthcare services may have profound consequences.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Explicability is a newly proposed principle within the particular AI context.

“It is to be noted here that explicability is a newly proposed principle, within the particular AI context, which has the same meaning outlined above in this paper [127].”

• cites [127] Preuer, 2019, Interpretable deep learning in drug discoveryNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

To achieve the potential of deep medicine, the scope of the to-be collected data should be wide enough to include speech, the intonation of voice, reaction times from keyboard use, GPS data, social media usage, and distinctive facial attributes in one's selfies.

“To achieve the potential of”

• cites [144] Association, 2001, World Medical Association Declaration of Helsinki. Ethical principles for medical research involving human subjects, Bull. World Health Organ.Not checked yet

  Not checked in the free run.

• cites [146] Berner, 2007Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The amount and type of needed data are increasing by time to the extent of seriously blurring the boundaries between what is medical and what is personal.

“The amount and type of the needed data are increasing by time to the extent of seriously blurring the boundaries between what is”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

EAMNet can detect diseases and show transparency by highlighting the affected area.

“The proposed architecture not only able to detect the diseases but also show transparency by highlighting the affected area detected by the system.”

• cites [123] Zafar, 2021, Deterministic local interpretable model-agnostic explanations for stable explainability, Mach. Learn. Knowl. Extract. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Bias in medical data is a major challenge and can eventually lead to algorithmic bias.

“In addition, bias in the data is also a major challenge to handle and that can eventually lead to algorithmic bias [178].”

• cites [178] Wang, 2019, Bias also matters: Bias attribution for deep neural network explanationNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Conducting proper analysis for all these data would necessitate giving access not only to one's physician but to many other experts in various areas.

“To make the situation more complex, conducting proper analysis for all these data would necessitate giving access not only to one's physician but to many other experts in various areas.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Various factors can put people's privacy at risk.

“Various factors can put people's privacy at risk, and we highlight here two of them: a) Unprotected Data Sharing:”

Probably not a substantial claim: may not pass S3 (load-bearing).

Traditional statistical performance metrics like the area under the curve may not be sufficient for evaluating models making clinical decisions.

“Moreover, it has been argued in the literature that traditional statistical performance metrics like the area under the curve may not be sufficient for evaluating the models making clinical decisions [176].”

• cites [176] Anwar, 2018, Medical image analysis using convolutional neural networks: a review, J. Med. Syst. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

ML techniques build their decisions on latent variables learned from data.

“ML techniques build their decisions on the latent variables which are learned from the data.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

IoT techniques are widely used in healthcare systems for real-time monitoring of critical patients.

“Techniques based on Internetof-Things (IoT) are widely used in healthcare systems for real-time monitoring of critical patients.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

With advanced technologies, the records and reports of patients are converted into electronic health records (EHR).

“With the advanced technologies, the records and reports of patients are converted into electronic health records (EHR).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

This ability leads to data breaching through tracking and monitoring of patient's routines which dishonors the patients' privacy.

“However, this ability leads to data breaching through tracking and monitoring of patient's routines which dishonors the patients' privacy.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Medical data is one of the most difficult data to handle due to its complex, multi-variate, and sometimes non-stationary and scarce nature.

“Medical data is one of the most difficult data to handle due to its complex, multi-variate, and sometimes non-stationary and scarce nature.”

May combine two claims.

Dependence of latent variables on each other can cause misleading patterns, and therefore ML-based decision-making will be misleading.

“The dependence of latent-variables on each other can cause misleading patterns and due to this issue, the ML-based decision-making will be misleading.”

May combine two claims.

Inherent bias in medical input data, such as class imbalance, can be reflected in model outputs and can influence model explanations.

“In addition, the inherent bias in the input (medical) data (e.g., class imbalance) can be reflected in the model's outputs i.e., model might prefer a specific class as compared to other classes and this bias might influence the explanations of the model [170].”

• cites [170] Klingler, 2017, Ethical issues in public health surveillance: a systematic qualitative review, BMC Public Health sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

These records are available online via the cloud servers.

“These records are available online via the cloud servers.”

May combine two claims.

Gradient-based explanations are dependent on the choice of reference point, and slight changes in the reference point significantly change the explanation, causing confusion and potentially misleading results.

“As these explanations are dependent on the choice of reference point and a slight change in the reference point of gradient will significantly change the explanation thus causing confusion that will eventually lead to misleading results or interpretation.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The blame game around accountability will foster epistemic vices such as dogmatism or gullibility.

“This blame game will eventually foster”

• cites [175] Weizenbaum, 1976Not checked yet

  Not checked in the free run.

May combine two claims.

Despite their significant performance, these models are black-box and lack theoretical understanding behind their decisions.

“However, despite their significant performance these models are black-box and lack theoretical understanding behind their decisions.”

May combine two claims.

Biases in medical data are hard to undo and their elimination has unintended consequences on results.

“These biases are hard to undo and their elimination have unintended consequences on the results [179].”

• cites [179] Weerts, 2019Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

It is well established that explanation of ML/DL techniques is required to gain the trust of clinicians in ML/DL empowered healthcare solutions.

“It is well established that the explanation of the ML/DL techniques is required to gain the trust of clinicians in ML/DL empowered healthcare solutions.”

• AssumptionRests on an assumption: Explanation of ML/DL techniques is required to gain the trust of clinicians in ML/DL empowered healthcare solutions.

Probably not a substantial claim: may not pass S3 (load-bearing).

A few studies suggested involving all types of stakeholders into the ML/DL method development process.

“In this regard, a few studies suggested involving all types of stakeholders into the ML/DL method development process that may include clinicians, policymakers, data scientists, ML researchers, and hospital staff, to name a few [182], [176].”

• cites [182], [176]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Visualization-based methods are widely applied to explain the decisions of ML/DL methods.

“In the literature, visualization-based methods are widely applied to explain the decisions of ML/DL methods.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Replacing human care with carebots will have social costs, for example exchanging feelings and emotions among humans will cease to be part of caregiving.

“Even if this automation of caregiving will not result in causing medical harm to the patient or job cuts in the healthcare staff, replacing human care will still have social costs, e.g., exchanging feelings and emotions among humans will cease to be part of caregiving [166].”

• cites [166] Atlam, 2019Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Explanations of DL models are model-specific in nature.

“As we discussed in Section III, that the explanations of the DL models are model-specific in nature, therefore, it is also required to develop inherently explainable and generalize explainable methods for the DL algorithms in future.”

May combine two claims.

The black-box nature of these models makes them susceptible to many vulnerabilities such as adversarial attacks, biased decisions, and inability to generalize out of distribution samples.

“Their black-box nature makes them susceptible to many vulnerabilities such as adversarial attacks, biased decisions, and being not able to generalize out of distribution samples, etc.”

May combine two claims.

ML/DL methods can be easily fooled and desired outcomes can be obtained.

“Over the past few years, it has been shown that ML/DL methods can be easily fooled and desired outcomes can be obtained [30], [31].”

• cites [30], [31]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Clinical impact of ML/DL advancements is only completely possible by overcoming challenges like adversarial ML attacks.

“Moreover, the clinical impact of ML/DL advancements is only completely possible by overcoming challenges like adversarial ML attacks.”

Probably not a substantial claim: may not pass S4 (based).

Developed explanation methods have unique challenges and limitations.

“However, the developed explanation methods have unique challenges and limitations associated with them, which are described below.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The critical nature of healthcare applications provides significant motivation for malicious actors to defame ML/DL-based systems and get desired outcomes.

“The critical nature of healthcare applications provides significant motivation for the malicious actors to defame the ML/DL-based system and to get the desired outcomes.”

May combine two claims.

A wide variety of adversarial ML attacks have already been proposed and research on developing respective defense methods is very limited.

“In the literature, a wide variety of adversarial ML attacks have been already proposed and the research on developing respective defense methods is very limited [73].”

• cites [73] Takabi, 2016, Privacy preserving multi-party machine learning with homomorphic encryptionNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Such an interdisciplinary development workforce will enable collaboration between knowledge experts and healthcare service providers that will eventually improve productivity and outcomes.

“Such an interdisciplinary development workforce will enable collaboration between the knowledge experts (i.e., clinicians and ML researchers) and healthcare service providers that will eventually improve productivity and outcomes.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Obtaining informed consent from the patient before exposing him/her to any medical intervention is one practical application of these principles.

“Obtaining an informed consent from the patient before exposing him/her to any medical intervention is one of the practical applications of these principles [135].”

• cites [135] Miles, 2005Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Modern bioethics have other approaches besides principlism.

“As mentioned above, modern bioethics have other approaches besides principlism.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Because of the significance of addressing data-related concerns, different countries have developed policies and laws.

“Because of the particular significance of addressing the data-related concerns, different countries have developed policies and laws [160].”

• cites [160] Haibe-Kains, 2020, Transparency and reproducibility in artificial intelligence, Nature sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

There is no doubt that deployment of ML for clinical practice will aid clinicians.

“There is no doubt that the deployment of ML for clinical practice will aid the clinicians.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Issues discussed within the care ethics approach go beyond solely measuring efficacy and safety of applications or calculating their possible health-related benefits and harms.

“The issues that can be discussed within this approach go beyond the question of solely measuring the efficacy and safety of certain applications or calculating their possible healthrelated benefits and harms.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The accountability situation becomes even more complex when all stakeholders in the loop are considered.

“This situation becomes even more complex when we consider all stakeholders in the loop.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Advancements in ML/DL techniques have a great potential to revolutionize healthcare.

“The advancements in ML/DL techniques have a great potential to revolutionize healthcare.”

• AssumptionRests on an assumption: Advancements in ML/DL techniques have a great potential to revolutionize healthcare.

May combine two claims.

Lack of ethical guidelines and fewer or no proper training of surveillance programs raises ethical concerns.

“Lack of ethical guidelines and fewer or no proper training of such surveillance programs raises ethical concerns [162].”

• cites [162] West, 2020Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Respect for persons and autonomy are among the widely-agreed upon principles in modern bioethics.

“As outlined above, the respect for persons and autonomy are among the widely-agreed upon principles in modern bioethics.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The care ethics approach focuses on the domain of intimate human relationships rather than the abstract application of rules.

“The care ethics approach, which focuses on the domain of intimate human relationships rather than the abstract application of rules [135], serves as a good example in this regard.”

• cites [135] Miles, 2005Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Due to the availability of implantable devices, it is now possible to monitor patients without their consent.

“Due to the availability of implantable devices, it is now possible to monitor the patients without their consent.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

HCI is a well-developed technique to empower users.

“For example, humancomputer interaction (HCI) is a well-developed technique to empower users.”

• AssumptionRests on an assumption: HCI is a well-developed technique to empower users.