Explainable AI in early autism detection: a literature review of interpretable machine learning approaches

Abstract

Autism spectrum disorder (ASD) is a neurodevelopmental condition especially in children with a strong hereditary basis, making its early diagnosis challenging. Early detection of ASD enables individualized treatment programs that can improve social interactions, cognitive development, and communication abilities, hence lowering the long-term difficulties linked to the disorder. Early detection helps in therapeutic interventions, which can help children acquire critical skills and lessen the intensity of symptoms. Despite their remarkable predictive power, machine learning models are frequently less accepted in crucial industries like healthcare because of their opaque character, which makes it challenging for practitioners to comprehend the decision-making process. Explainable AI (XAI), an extension to AI, has emerged due to issues like trust, accountability, and transparency caused by the opaque nature of AI models, especially deep learning. XAI aims to make AI’s decision-making processes easier to understand and more reliable. The present study delves into the extensive applications of XAI in diverse fields including healthcare, emphasizing its significance in guaranteeing an ethical and dependable implementation of AI. The article goes into additional detail in a specialized assessment of AI and XAI applications in research on ASD, showing how XAI can offer vital insights into identifying, diagnosing, and treating autism.

Introduction

AI has been applied in nearly every field, including manufacturing industry, legal systems, healthcare, security, and education [1]. Alan Turing, in his 1950 seminal work'Computing Machinery and Intelligence,'first came up with the term machine intelligence and also came up with the Turing Test as a measure to test if a machine is capable of demonstrating intelligent behavior equivalent to that of a human [2]. AI has advanced significantly and is increasingly integrated into various aspects of daily life, offering capabilities that can complement and, in some cases, match human performance in specific tasks [3]. AI has very efficiently found its place in almost every domain like disease prediction, recommendation engines, automated cars, automated robots, agriculture, defense, smart assistants and so on [4–8]. Thus, it is evident that it has influenced every aspect of our daily lives as we unknowingly use underlying AI tools to recommend friends on social media platforms, choose movies to watch, and make purchases on e-commerce websites.

Despite having numerous applications, predictor ML based models present a considerable problem due to their opaque nature. Even while these algorithms are easy to use and accurate in producing findings and predictions, they frequently act as “black boxes,” making it difficult to interpret them. This lack of interpretability and transparency raises concerns, particularly when crucial decisions need to be made in domains like healthcare, defense, legal systems and finance.

The need for more transparency in interpreting how this AI model arrives at their predictions or decisions creates a hindrance and limitations in its use in critical systems. To combat this, XAI has recently gained the considerable attention of researchers and developers. The expanding research conveys that the attention of XAI is not only limited to academics but also extends to a wide array of domains where the understandability of AI models stands on priority to agree or disagree with the decision made [9].

The main goals of XAI are to: (1) create highly interpretable models that explain the predictions without affecting learning performance, thereby delivering high levels of confidence and prediction accuracy; and (2) make it possible for humans to efficiently manage and accept the next generation of artificially intelligent collaborators. XAI due its applicability of converting the opaque nature of black box models into transparent ones, have found its applicability in a variety of critical domains.

ASD—a complex neurodevelopmental condition is characterized by challenges like social connection, communication, and repetitive behaviors. Early diagnosis of ASD is crucial, as timely interventions have been shown to significantly improve developmental outcomes, enhancing social, cognitive, and adaptive skills. However, diagnosing ASD, particularly at an early stage, remains challenging due to its heterogeneous nature and the reliance on behavioral assessments, which are often subjective and time-consuming.

The characteristics of any neurological condition typically first seen in early life are disturbances to social interaction, communication, & behavior patterns. The complex etiology of Neuro Developmental Disorders (NDDs) combines several environmental and genetic factors [10]. Early identification of any disorder is important as it enables early therapies to enhance the developmental outcomes and quality of life for such people. However, early diagnosis can be difficult, given the variety of symptoms and differences in how they appear. Traditional methods of diagnosis, though useful, mainly depend on behavioral aspects and observations, which could lead to a delay in diagnosis. This delay emphasizes the critical need for reliable techniques to detect it as early as possible, preferably by using biological markers such as genetic data analysis and neuroimaging scans. Given that several genetic mutations and variants have been associated with the development of neurological disorders, the scope of genomic studies has led to interesting avenues in research. Due to the advancements in genome and exome sequencing, which have become more readily available, researchers may now examine a large range of genetic data to find possible indicators and risk factors. However, large volumes of genomic data make it time-consuming and computationally demanding to isolate particular markers relevant to study from massive amounts of genetic databases. Moreover, the complex character of genetic data sometimes calls for advanced analytical techniques to control big datasets and discover significant trends. In such situations, XAI methods present a revolutionary solution [11].

XAI methods can aid in making predictions made by AI models for early detection more interpretable and transparent. Traditional machine learning (ML) methods of prediction are considered “black-box” models. XAI facilitates bringing clarity to the decision-making process of such opaque models, thus making them more reliable for researchers and clinicians. Transparency is crucial in life-critical application fields like medical data analysis, where knowledge of the"why"behind a prediction is essential to ensure trustworthiness and reliability in the outcome.

XAI methods thus enable researchers not only to predict disorders under specific parameters but also to understand how the specific factors and pathways are affecting these predictions. The specific nature of results will enable the counselor and healthcare providers to advise and make decisions that are based on an informed understanding, which is an essential factor between high-order data and practical clinical applications. Integrating XAI with medical data analysis also helps with precision medicine in treatment. This tailored technique may help diagnose more precisely and offer better prospects for targeted and successful treatments according to an individual's genetic makeup. Such disorders vary widely, and thus, these individualized approaches may be revolutionary.

The work intends to provide interpretable insights that support therapeutic decision-making by means of XAI approaches to analyze complex data, discover major factors linked with it, and generate a framework. The results of this study could offer a creative, statistically based method for early disorder diagnosis, enabling earlier intervention and better long-term results for people with the same.

This paper presents a thorough literature review of recent XAI research in the healthcare industry so that researchers may use it for their research. The paper is structured as follows: Sect. 2 introduces the key XAI concepts, terms, techniques and overview of the categories of XAI methods. Section 3 explores the applications of XAI in various critical domains. Section 4 delves into the advancements and research in XAI specifically within the healthcare industry. Section 5 is dedicated to a focused survey on the application of XAI in autism research, highlighting the significant work recently done in this area. Further sections explain the research gaps in the existing survey as to how we can improve the transparency, interpretability, and reliability of AI-driven diagnostic models by integrating XAI tools into the early autism detection framework.

XAI concepts and terminologies

Integrating XAI into ML models

For AI-driven decision-making processes to be more transparent, trustworthy, and interpretable, XAI module must be integrated into ML models. By enabling users to understand and analyze the generation of outputs by ML models, XAI approaches transform traditionally opaque “black-box” models into more understandable systems.

Figure 1 shows the role of the XAI module in the complete process of deployment of ML models. As it is clear, by offering insights into the decision-making processes, XAI supports the identification of significant characteristics, model debugging, and algorithm selection during the model building and training phases. XAI ensures that models function properly and equitably during the assessment and validation phase by providing thorough explanations of model behavior, assisting in the discovery of bias, and fostering stakeholder trust. When models are deployed, XAI guarantees regulatory compliance, promotes user adoption via transparency, and helps with risk management by pointing out any problems. When used in an operational setting, XAI helps decision-making by offering clear insights, making it easier to track the performance of the model continuously, and incorporating user feedback for continual improvement. After deployment, XAI supports incremental improvements to AI systems, frequent compliance audits, and retrospective study of model decisions. While techniques such as Grad-CAM and Saliency Maps emphasize salient regions in images, LIME and SHAP algorithms explain individual predictions and feature contributions. Interpretable standards for more complicated models are decision trees and rule-based models. By including XAI modules at every stage of the ML lifecycle, you can improve transparency, accountability, and user engagement while guaranteeing that AI systems are fair, effective, and compliant with user and regulatory requirements.

Fig. 1.

Role of XAI in ML Models

As shown in Fig. 2, the “black box” aspect of many complex AI models is addressed with XAI, allowing for transparency and an understanding of the decision-making process. Establishing trust with users and stakeholders is crucial to guarantee that AI systems are not just efficient but also impartial and equitable. Another important consideration is regulatory compliance, since many sectors demand that automated judgments be justified and auditable. Additionally, by providing insights into feature relevance and decision routes, XAI improves model construction and debugging, resulting in more durable and dependable AI systems.

Fig. 2.

Need of XAI

AI is swiftly and broadly being implemented into daily life as we reach Industry 4.0, which is speeding up the shift to an increasingly algorithmic society. However, despite these remarkable developments, a significant obstacle to the widespread use of AI-based systems is their frequent opacity. These systems are designed in a way that makes them tough to explain, but it makes strong predictions possible. XAI has recently been the subject of discussion due to this concern. There is a lot of promise for this research to improve the reliability and transparency of AI-based systems. It is recognized as being critically crucial for AI to develop in the future in a steady and unrestricted manner [9].

The concept of XAI is not new. The first research on XAI was done forty years ago in a paper where some expert systems used the applied rules to explain their outcomes. Since the start of AI research, scientists have maintained that intelligent systems ought to be responsible for explaining AI outcomes, particularly in decision-making. When a rule-based expert system declines a credit card payment, it ought to provide an explanation for its decision. Expert systems contain knowledge and rules that are simple for humans to comprehend and interpret since they were created and established by human experts [12].

A DNN model can be characterized using two terms- interpretability and explainability. The confidence of developers increases because interpretability lets them dive deep into the model’s decision-making process. Rather than offering a mere forecast, the interpretation method offers an interface that furnishes supplementary details or elucidations that are required for deciphering the fundamental operations of an AI system [13].

On the other hand, explainability informs the user of the DNN’s decision, increasing user confidence that the AI is drawing fair, fact-based judgments. Explainability in the context of AI ideas refers to the model’s functional knowledge, which tries to describe the model's behavior in a black-box environment [14].

Building human-interpretable models is the primary goal of XAI, especially for applications in critical industries like banking, healthcare, and the military. Although domain specialists need help to solve problems more quickly, they also need significant results in order to understand and be confident in the answers they receive. Domain experts can benefit from examining appropriate outputs, but developers can also benefit from this process if the outputs turn out to be erroneous, as it compels them to investigate the system inside. AI techniques enable (i) the assessment of current knowledge, (ii) the advancement of knowledge, and (iii) the formation of new conjectures and assumptions [15].

As highlighted by [16], there is no one-size-fits-all XAI solution; rather, the selection of an explainability technique should be tailored to fit the particular context, model design, data type, and user requirements. This supports the need for context-sensitive XAI choice in safety–critical areas like healthcare and ASD diagnosis, where explanations need to be technically sound as well as clinically significant.

XAI methodologies

Techniques for explainability come in various forms and dimensions, as shown in Fig. 3. They fall into four main categories: methodology-based, complexity-based, model-based, and scope-based. Although there are numerous methods for assessing explainability, they will all be covered in length in the paragraphs that follow.

Fig. 3.

XAI Methodologies

Methodology based explanation

Depending on the methodology used, XAI core algorithms can be divided as either perturbation-based or backpropagation-based. It is possible to backpropagate an important signal from the output to the input using techniques based on backpropagation. This starts with the network's output and increases the weight of each intermediate value determined by the forward pass. A gradient function determines how the network output relates to each intermediate parameter, allowing the weights of each parameter to be updated and the output to be aligned with the ground truth. Therefore, another term for these methods is gradient-based [17].

Conversely, perturbation-based algorithms change the feature set of an input instance and examine the effects of these modifications on the network output by employing occlusion, partially replacing features through filling operations or generative algorithms, masking, conditional sampling, and other strategies. In this situation, one forward pass is sufficient to understand how the perturbed component in the input instance contributes to the network output, negating the necessity for backpropagating gradients [18].

Model based explanation

Interpretability methodologies that are currently in use can also be categorized using model-agnostic or model-specific methods. As the name suggests, a model-specific approach is limited to specific types of models. Intrinsic methods are model-specific by definition [19]. Model-agnostic techniques, on the other hand, don’t care which kind of ML model is applied. There is a recent surge in interest in model-agnostic interpretability techniques because they are model-free. Post-hoc interpretability is provided by model-agnostic techniques, which are frequently applied to interpret ANNs as local or global explainers [20].

Scope based explanation

A popular technique for figuring out how model outputs relate to inputs that either display the behavior of the entire model or just one prediction is feature significance analysis. A local or global approach may be used for the type of analysis, depending on how significant the feature is. Local explainers simply provide an explanation for a single choice or situation. This suggests that the decisions they make are restricted to a particular scenario and one possible explanation. A classic illustration of a local explanation is LIME. Global explainers, on the other hand, offer an explanation for the entire dataset. These hypotheses still support overall observations. Localized explanations may also be provided by some global explainers, though. For instance, SHAP can offer both global and local explanations [21].

Complexity based explanation

The extent of interpretability is closely correlated with the ML model’s complexity. In general, a model becomes harder to comprehend and describe the more complex it is. Depending on when interpretability is attained, interpretable ML algorithms fall into one of two categories: intrinsic or post-hoc interpretable. By developing models that are self-explanatory and have interpretability built right in, intrinsic interpretability can be achieved. Stated otherwise, the structure of intrinsic interpretable models is straightforward. However, accuracy is frequently sacrificed in favor of the models’ readability and simplicity [22].

Another possible approach would be to create a high-complexity, high-accuracy model and then use an alternative set of techniques to provide the required explanations without having to comprehend how the original model functions. This family of approaches provides post-hoc explanations. To give consumers an explanation, post-hoc interpretability entails creating a second model, which is typically a proxy of the original model [14].

XAI techniques summary

Table 1 briefly summarizes a few of the techniques used for explanations in ML models and the type they belong to.

Table 1.

Summary of XAI Techniques

Technique Name	References	Gradient/Perturbation	Model Specific/Model Agnostic	Local/Global	Intrinsic/Post-Hoc
LIME	[23–26]	Perturbation	Model Agnostic	Local	Post Hoc
SHAP	[27–29]	Perturbation	Model Agnostic	Both	Post Hoc
LRP	[30]	Gradient	Model Specific	Local	Intrinsic
CAM	[31, 32]	Gradient	Model Specific	Local	Post Hoc
Grad-CAM	[33]	Gradient	Model Specific	Local	Post Hoc
Saliency Map	[34–37]	Gradient	Model Specific	Local	Post Hoc

XAI applications in discrete domains

XAI in cyber security

Cybersecurity is a field in computer science that guards against theft, damage, and unwanted access to computer systems, networks, and data. The surge in internet use across the globe has enormously increased the threats of cyber-attacks like malware, phishing, Denial of Service (DoS) attacks, and Brute Force attacks. Many techniques, processes, and practices are designed and used to protect the IT infrastructure and ensure data confidentiality, integrity, and availability.

Initially, the rule-based and conventional ML approaches came into the picture for combating cyber-attacks. As new networking paradigms like the IoT, cloud computing, and fog computing, traditional algorithms have started to show limitations in processing and protecting massive amounts of data, and computing costs have risen. Few researchers have contributed to these concerns using AI, ML and DL approaches. However, the researchers faced specific upcoming challenges due to certain limitations of AI models. One such severe challenge was that of the opaque behavior of AI models. The problem was related to the rationality and justifiability of the decisions generated by the AI models that were difficult for people to understand and decode. Various approaches of XAI can be very effectively applied in combating cyber security attacks like malware, DoS, spam, and frauds [38].

XAI can address critical challenges like trust and accuracy in cyber security. Addressing privacy concerns and clear explanations in AI-based intrusion detection systems is also essential [39].

SHAP based XAI functions provide insights into feature importance for each decision, enhancing the understanding of decision-making processes for security administrators and cyber security analysts, fostering trust in the AI-powered IDS [40].

XAI in legal systems

Incorporating AI in the legal industry has brought a massive transformation by offering several benefits, such as speedy analysis of raw data, which supports decision-making and has increased efficiency in various tasks for legal professionals. Whatever advantages AI tools provide, they cannot replace lawyers as they lack critical thinking, creativity, emotional intelligence, and the ability to navigate legal proceedings. Despite its advantages, AI in law faces challenges, including vulnerabilities to threats and the need for a comprehensive legal framework to regulate its behavior and protect user data [41].

A case study of Article 6 cases at the European Court of Human Rights presented AI-based tools to enhance the justifiability of decisions by predicting the outcomes of cases. The tools prioritized interpretability by explaining the predictions and achieved 97% accuracy in predictions but showed slower case processing [42].

AI tools greatly facilitate and accelerate decision-making, but AI and law cannot be considered one-size-fits-all. Depending on the type of legal field, different logical reasoning and levels of explainability are required. There is much scope in researching various explainable tools for various legal fields [43].

“Intelligent Justice” can be achieved by carefully combining AI, Natural Language Processing (NLP), XAI, blockchain, and privacy techniques. The combined framework improves efficiency and accuracy and incorporates transparency through explainability and privacy in the legal systems. When users are able to evaluate the recommendations and judgments and comprehend the rationale behind them, their level of happiness will rise [44].

XAI in finance

XAI plays a remarkable role in the finance sector as it provides excellent transparency and interpretation of complex decision-making processes like Credit scoring, fault detection, ethical and fair AI practices, market analysis, and compliance and regulations. A thorough examination of the XAI techniques applied in the financial sector, with an emphasis on assessing their effectiveness across a range of activities and domains is given in [45].

A case study that used AI and ML models to anticipate financial hardship from 2010 to 2021 in a few companies in Vietnam identified algorithms such as extreme Gradient Boost (XGBoost) and Random Forest (RF) along with SHAP, which revealed significant results and predictions for the same. While shedding light on using XAI in economic decision-making, the study acknowledges limitations like a low sample size. It suggests future research will expand samples and consider additional behavioral features beyond financial indicators in Vietnam's unique market context [46].

With the advent of digital payment systems, the responsibility of payment providers to reduce the number of fraudulent transactions has become even more crucial. The growing issue of fraud can be controlled to some extent by applying XAI models as they not only predict the fraudulent transactions but also can provide explanations and help in combating this fraud [47].

Similarly, XAI can be used to develop unbiased and fair credit scoring models. These models help lenders decide whether to accept or reject loan applications from applicants. For this, the model predicts whether the applicant can repay the amount based on the data the applicant provided. A model consisting of the XGBoost algorithm for classification along with several XAI techniques like SHAP along with Generalized Relevance Integrated Propagation (GRIP), Anchors, and ProtoDash provided three types of explanations—local-instance based, local-feature based and global, also with the predictions. The financial experts found the explanations to be beneficial and that it can be implemented in the system of various banks [48].

XAI technique, along with the classifier, is not only used to recognize causes of financial distress but it can also be used to improve the variable selection operation and the testing accuracy to 96%, which is better than other classifiers [49].

XAI in education and training

AI has been utilized in the Artificial Intelligence Education System (AIED) as it automates several traditional education approaches. It can be used in various aspects of education like personalized teaching, efficient resource allocation, enhancing outcome-based education, and formulating educational policies.

AI technology is anticipated to grow quickly in the education sector. From 2018 to 2022, the US industry is seen to account for 48% of the global AI market growth. AI technologies have the potential to expand learning possibilities, personalize experiences, and enhance learning tactics. In the past, academics have suggested that AI robots may replace instructors in future. Due to the advent of AIED, this is now gradually happening. Happy Numbers, the first AI based teaching robot, has worked in US classes. Collaborations across disciplines and a critical knowledge of ethics are necessary for integrating AI in education. The research in similar domains will result in practical guidelines and new teaching methods [50].

The introduction of XAI has been a boon as it addresses the opaque nature of AI models, which was a lacuna due to increasing demand for AI in education and training purposes. Various model-specific explainability approaches have been mentioned to incorporate interpretability into AI models in education [51].

XAI-ED is a framework that addresses six crucial aspects for incorporating explainability to analyze, design, and develop educational related AI tools. These dimensions encompass stakeholders involved, anticipated benefits, strategies for conveying explanations, prevalent categories of ML models, human-centric designs for AI interfaces, and prevailing challenges associated with offering explanations in educational contexts. The application of XAI-ED is exemplified through four in-depth case studies, each demonstrating its relevance and effectiveness in diverse educational AI tool scenarios [52].

XAI in healthcare

A crucial development in healthcare that guarantees openness and confidence in AI-driven decision-making procedures is XAI as illustrated in Fig. 4. XAI helps close the communication gap between medical practitioners and sophisticated algorithms by offering insights into how AI models arrive at certain diagnosis or treatment suggestions in a transparent and intelligible manner.

Fig. 4.

Applications of XAI in Healthcare

In order to assess AI outputs and make sure they correspond with clinical knowledge and patient details, practitioners must have this openness. A detailed review of taxonomy and analysis for recently proposed existing AI based medical image processing approaches in healthcare is given in [53].

Furthermore, as patients are more likely to accept and follow treatment programs that they understand, XAI improves patient trust. Furthermore, when decision-making is open, regulatory agencies are better able to evaluate the usefulness and safety of AI applications in the healthcare industry. In the end, XAI creates a cooperative setting where AI functions as an auxiliary tool, providing doctors with trustworthy, understandable, interpretable insights.

Recently, a lot of work has been done in this area, considering the growing need for trustworthiness and robustness in using machine-predicted results in healthcare. Figure 5 shows the number of publications on Scopus and Web of Science Database over the years with the search query—“Explainable AI in Healthcare”, “XAI in Healthcare”.

Fig. 5.

No. of publications vs. years as per a Scopus and b WoS

In an analysis that illustrated current trends in AI research connected to health, showing that the growth rate of publications on AI in the healthcare industry has been steadily increasing in the last several years [54].

Healthcare data processing must be clear, according to current privacy rules like GDPR, HIPAA, and PIPL. This implies that AI algorithms that use this data must also be transparent and understandable. AI explainability will probably be subject to even stricter privacy rules in the future. However, explaining AI is challenging and will only worsen as AI algorithms become more complicated. Due to their growing complexity, it may become nearly impossible for patients and clinicians, who are the end consumers of healthcare, to comprehend and trust the algorithms. The primary concern of end users, particularly patients, is to ensure that the algorithm’s predictions are reliable. This can be achieved by displaying previous accurate forecasts to them. The harmony between explainability and performance or accuracy is another crucial concern. Accuracy and performance are important because they may mean the difference between life and death, especially in the healthcare industry [55].

An extraordinary period of advancement is being brought in by AI-driven discoveries in the formulation of treatment protocols, medication manufacturing, and the identification of novel biomarkers. Data integrity, universal application, and ethical issues are still significant concerns, though. These issues need to be resolved, and a lot of research needs to be done, in order to justify AI’s function in Breast Cancer detection and therapy [56].

In [57], the author has proposed a methodology employing various ML algorithms for early detection of Ovarian Cancer. By using Support Vector Machines (SVM), 89% accuracy is obtained for the basic model and 86% accuracy for the model after stacking multiple ensemble learning techniques. Complex ML algorithms can benefit from deeper insights into their decision-making through the use of XAI, which increases the algorithms’ applicability. This study aims to present the best practices for combining AI and ML with biomarker evaluation. The research focused on developing and assessing Shapley values-based classifiers and visualizing the outcomes. The study offers a promising method for the early diagnosis of ovarian cancer, which benefits women’s health and the area of oncology.

In a study by [58], the author evaluated the performance of Saliency Maps that are widely used in detecting and explaining breast cancer based on mammograms. On a balanced mammography dataset of women with 1496 cancer-positive and negative images from various centers, three radiologists created ground-truth boxes. Using mediolateral oblique (MLO) and craniocaudal (CC) pictures, a modified, pre-trained DL model was used to detect breast cancer. The Grad-CAM, Grad-CAM +  +, and Eigen-CAM saliency XAI approaches were assessed. The model’s recall, precision, accuracy, and F1-Score in identifying cancer in the testing set were around 68%, 87%, 81%, and 0.77, respectively.

The new interpretable hybrid deep learning-based lung cancer detection method called “DeepXplainer” is presented in [59] paper along with an explanation of the predictions made. This method is predicated on XGBoost and CNN. After “DeepXplainer” has automatically learned the features of the input using its several convolutional layers, XGBoost is employed for class label prediction. SHAP, is used to provide explanations or assess the explainability of the forecasts.

In [60], the author has developed a system called “Explainer” that detects the presence of lesions in thyroid glands by producing the heatmaps for explanation of predictions. The Explainer provides doctors with a tool to understand the foundation of AI forecasts and assess their dependability, which may help to open the “black box” of AI in medical imaging.

In [61], XAI was used on XGBoost ML models. The algorithm was developed using a binary classification dataset that included Peripheral Blood Mononuclear Cell (PBMC) expression data from about 250 patients with breast cancer and 200 healthy women. Finding possible breast cancer diagnostic biomarkers was the aim of this investigation. Ten significant genes linked to the development of breast cancer were shown to be effective prospective biomarkers after successfully incorporating SHAP values into the XGBoost model.

In order to accurately diagnose skin cancer, in [62], the research modifies the pre-trained MobileNetV2 and DenseNet201 deep learning models by including three extra convolution layers. The enhanced models outperform the pre-trained MobileNetV2 and DenseNet201 models, according to a thorough examination. The proposed method allows for the detection of both benign and malignant classes. The results demonstrate that the proposed Modified DenseNet201 model achieves 95.50% accuracy and great performance when compared to other techniques. Additionally, the Grad-CAM approach is employed to efficiently depict the CNN results. [63] Presents the NeuroXAI framework for XAI of DL networks. With the use of seven cutting-edge explanation techniques and visualization maps like Vanilla gradient, Guided backpropagation, integrated gradients, guided integrated gradients, SmoothGrad, Grad-CAM and guided Grad-CAM, NeuroXAI contributes to the transparency of deep learning models. Two of the most extensively studied issues in brain imaging analysis—image classification and segmentation utilizing magnetic resonance (MR) technology—have been tackled with NeuroXAI. For both applications, visual attention maps of several XAI techniques have been created and compared.

Distinct ML methods (LR, ANN, RF and XGB) were used in [64] in order to construct the best predictive model for Hypertension (HTN) categorization. According to the overall experimental results, the XGB model is the most suitable model among the four models for predicting patients at risk of developing HTN. The main risk factors of developing HTN are—Age, weight, fat, income, BMI, diabetes, salt, HHTN, drinking, and smoking according to the research. As a result, the suggested integrating system may be easily applied as a helpful tool in clinical settings to precisely identify individuals at an early risk of HTN. The contributions that XAI models make are highly valuable and drive the process toward more sophisticated approaches and measures, even if XAI evaluation still has to be refined and validated on various datasets, machine learning models, and XAI techniques [65].

[66] Experimented with the performance of four classification algorithms: Light Gradient Boosting Machine (LGBM), GB, RF, and ABC. LGBM performed the best of these, averaging an amazing 99.33% for training accuracy. According to the findings, RF, GB, and LGBM showed similar accuracy, recall, and F1 scores; they could also distinguish “Normal” cases with high precision and show strong recall for “Attack” cases. By comparison, ABC showed the greatest misclassification rate and lagged in every measured feature. Additionally, a LIME analysis of feature selection from XAI algorithms revealed that “troponin” and “kcm” had a greater significance in predicting the sample as an assault.

In [67], a unique AI method is developed using the Explainable Boosting Machine (EBM), a tree-based algorithm, to anticipate Anigma Pectoris (AP) in women. EBM is a ML technique that combines GB accuracy and flexibility with the interpretability of linear models. EBM is used to identify the most important factors for AP prediction from a sample of 200 female patients—100 with AP and 100 without. Next, the effectiveness of EBM with various AI techniques, including Light Gradient Boosting Machine (LightGBM), Adaptive Boosting (AdaBoost), XGBoost, LR, and Categorical Boosting (CatBoost) was compared. It was discovered that EBM, with an accuracy of 0.925 and a Youden’s index of 0.960, was the most balanced and accurate method for predicting AP.

In [68], a dataset comprising 1,120 ECG recordings from healthy individuals ranging in age is analyzed using two distinct models: XResNet50 and XGBoost. The XResNet50 model uses raw ECG data, but the XGBoost model accepts both long-range and short-range ECG features as input. Both models were trained to predict the age of a healthy individual from their 1-lead ECG, and they showed competitive performance (macro-AUCs of 0.73 and 0.77, respectively). Following training, both models were examined using XAI techniques, and the most crucial ECG features for age classification were identified by analyzing the XGBoost model using SHAP.

The work in [69] presents the development of an autonomous diabetes prediction system employing different ML algorithms and a private dataset of female patients in Bangladesh. The mutual information feature selection algorithm has been applied in this work. A semi-supervised model featuring high GB has been utilized to forecast the insulin properties of the private dataset. The class imbalance issue has been addressed using SMOTE, Adaptive Synthetic Sampling Approach techniques (ADASYN). The authors employed a variety of ensemble techniques together with ML classification algorithms, such as decision trees, SVM, RF, LR, KNN, and others, to ascertain which algorithm yields the most accurate predictions. After training and testing all the classification models, the recommended technique yielded the best result in the XGBoost classifier with the ADASYN approach, achieving 80 percent accuracy. Using the LIME and SHAP frameworks, the XAI technique is used to comprehend how the model predicts the outcome.

In [70], an ensemble classifier has been used with six ML algorithms—RF, LR, artificial neural network (ANN), SVM, AdaBoost, and XGBoost—to detect diabetes. To aid doctors in comprehending the model predictions, SHAP have been used to generate global and local explanations for each ML model. These explanations are displayed in several graph formats. Using a five-fold cross-validation (CV) approach, the created weighted ensemble model yielded a balanced accuracy of 90% with an F1 score of 89%. The dataset’s classes were balanced using the SMOTE technique, and the missing values were imputed using the median values.

The study in [71] will act as a catalyst for converting important research in Alzheimer’s into significant therapeutic outcomes since it offers a unique comparison of various AI techniques using the same benchmark datasets. In [72], a framework is developed for classifying Alzheimer’s disease (AD) using multimodal data, which includes genetic information, MRI images, and tabular data. The pre-processing stage generates a knowledge graph using tabular data and MRI pictures. Then, region-based CNN is applied for image-to-knowledge graph development, and graph neural networks are used for knowledge graph creation. Furthermore, multiple XAI methods are employed to decipher and explain the prediction results obtained from multimodal data. The layer-wise results in the MRI pictures were explained by layer-wise relevance propagation. Additionally, interpretable model-agnostic submodular select local explanations were included to interpret the decision-making process in light of the given tabular data. The Submodular pick LIME (SP-LIME), LRP, and GGT methodologies were employed to ensure dependable interpretation. The characteristics of AD, such as age, mental state, and clinical dementia, are interpreted by SP-LIME. LRP identifies significant brain Regions of interest (ROIs) for AD patients, which is important for analyzing the severity of the disease. By looking at these ROIs, it is simpler to determine which brain regions are in charge of particular kinds of ADs. Experts and physicians can easily identify the ROIs that are important for AD. Furthermore, we identified the biology of AD patients using GGT. The results of this biological interpretation will aid specialists in comprehending the genes that significantly contribute to AD. Medication prescriptions and treatment depend on gene analysis.

The application of clustering algorithms after explainable classifiers is still mainly unexplored, although it offers potential for new insights into illnesses. Specifically, it allows for a better understanding of dynamics compared to classifiers and more selective pattern recognition compared to clustering techniques. In [73], a similar approach is used. A method for using whole-brain dynamic functional network connectivity (dFNC) data is proposed to construct an explainable DL model that distinguishes people with Schizophrenia (SZ) and Healthy Controls (HCs). Then, the ensuing explanations are grouped into clusters to find distinct brain activity states. Specifically, it is found that schizophrenia has a widespread impact on how the subcortical, sensory (i.e., visual and somatosensory), and cerebellar networks interact. Several additional characteristics are also extracted to measure various elements of the classifier explanations, and the findings suggest that there may be a temporal localization of the impacts of SZ and a reduction in total brain activity variability in people with SZ.

In [74], the study’s objectives were to create trustworthy ML prediction models for stroke illness, handle a common severe class imbalance issue that arises from the stroke patients’ class being noticeably smaller than the healthy class, and analyze the model’s output to comprehend the decision-making process. In a comparative examination with six popular classifiers, the efficacy of the suggested ML approach was examined in relation to metrics that pertain to prediction accuracy and generalization capabilities. At around 19%, the Multi-Layer Perceptron (MLP) classifier performed the best overall in terms of false-negative rates. SHAP were employed to investigate the impact of the risk factors on the prediction’s result.

[75] Classifies the ischemic stroke group and the healthy control group using ML models in order to predict acute stroke in active stages. Additionally, the model's behavior was elucidated, and the key components of stroke prediction models were identified by using XAI tools, including Eli5 and LIME. In this study, 75 healthy persons without a history of neurological conditions, as well as 48 patients who were hospitalized for an acute ischemic stroke, were looked into. Three months after the onset of symptoms of an ischemic stroke, an Electroencephalography (EEG) was recorded utilizing frontal, central, temporal, and occipital cortical electrodes (Fz, C1, T7, Oz). EEG data were gathered while the subject was actively engaged in walking, working, and reading. The Adaptive GB models in the ML approach’s results showed about 80% accuracy in classifying the stroke group and the control group. Eli5 and LIME were then applied to explain the behavior of stroke prediction model, which were also used to interpret the model locally surrounding the prediction. The stroke-prediction XAI model is anticipated to aid in post-stroke therapy and recovery and assist healthcare professionals in making more explicable diagnostic judgments based on the outcomes of this XAI research.

About autism

The complicated neurological disorder known as ASD is characterized by issues affecting social interaction, communication, and routine tasks. Early identification and diagnosis of ASD are essential for starting effective interventions on time, which can significantly enhance the quality of life for those with the illness. Using ML approaches to improve early detection of ASD has come a long way in the last several years. Researchers have used numerous machine learning algorithms to examine various datasets, including behavioral patterns, social interaction cues, genetic data, and brain imaging data. Predictive models that detect ASD early have been developed, opening the door to earlier intervention and better results. Figure 6 shows the major types of NDD in children: ASD (Autism Spectrum Disorder), ADHD (Attention Deficit Hyperactive Disorder), Dyslexia, and Epilepsy. According to the World Health Organization (WHO), approximately 10–15% of children globally are affected by NDDs, including ASD, ADHD, and Intellectual Disability. As of 2024, the Centers for Disease Control and Prevention (CDC) reported that in the United States, 1 in 36 children are diagnosed with ASD. Globally, the prevalence of ADHD in children is estimated to be about 5–7%. In India, studies estimate the prevalence to be 1.6 to 17.9%, depending on the region and diagnostic criteria used.

Fig. 6.

Major Types of NDDs in children

In the further section, we emphasize ASD among the other kinds of NDDs because of its intricate genetic foundations, increasing incidence throughout the world, and the substantial influence that early diagnosis can have on long-term developmental outcomes.

AI and XAI work in autism detection

XAI and AI have an important role in the detection and diagnosis of ASD with cutting-edge methodologies for early detection and individualized treatment. ML algorithms such as support vector machines (SVM), random forests, and deep models, have been extensively used to analyze a wide range of data types such as neuroimaging, genetic, behavioral, and clinical data. These AI models have the ability to detect patterns and features that are hard for human clinicians to detect, resulting in more accurate and timely diagnoses. Yet, the black-box nature of most AI models poses problems in clinical environments, where interpretability is critical to trust and adoption. This is where Explainable AI (XAI) steps in, offering techniques such as LIME, SHAP, and attention mechanisms that make these models more transparent by interpreting their decisions in a human-readable manner. In combination, AI and XAI are a formidable pair for enhancing ASD detection, with the promise of earlier intervention and improved outcomes for autistic individuals.

In [76], the authors have provided a comprehensive overview of the current trends and future directions in machine learning techniques for diagnosing ASD. They highlight the significant potential of various ML models, such as SVMs, CNNs, and deep learning architectures, in improving diagnostic accuracy.

Table 2 presents a survey on the recent work done in detecting ASD with the use of AI and XAI techniques.

Table 2.

Related work in the detection of ASD

Refs.	ML Methods	XAI	Data Type/Source	Age Group/Size	Key Findings	Limitations
[77]	BERT + Multi-Head CNN	None	ABIDE-I: fMRI + clinical data	Leave-one-site-out (ABIDE-I)	Achieved 93.4% accuracy using multimodal fusion of imaging and meta-features	No explainability layer; generalizability to other datasets untested
[78]	Bi-LSTM + White Shark Optimization (WSO)	None	Autism screening datasets: Toddlers, Children, Adults	3 datasets; leave-one-dataset-out validation	Achieved high accuracy (97.6%, 96.2%, 96.4%) across datasets using WSO for feature selection and Bi-LSTM for classification	No explainability layer; generalizability to other datasets untested
[79]	Bat-PSO-LSTM, Feature Fusion	Not mentioned	Clinical/Autism-related	Various age groups	Improved accuracy in ASD diagnosis with feature fusion and Bat-PSO-LSTM	Limited dataset size and generalizability details
[80]	LR, SVM, RF, NN, DT, kNN	2TLFFDOSM (Fuzzy MCDM)	Medical + Socio-demo (PCA-fused)	1296 samples (all triage levels)	LR showed top robustness across 18 criteria	No per-patient XAI; model ranking only
[81]	LR, SVM, RF, k-NN, NB, DT, NN	LIME, Fuzzy MCDM (2TLFFDOSM)	Medical + sociodemographic	538 ASD cases (all ages)	Fuzzy-XAI framework for ASD triage with interpretable feature-based decisions	Limited external validation; LIME instability noted
[82]	Xception CNN	Not mentioned	Neuroimaging data	Various age groups	Xception model for ASD diagnosis using neuroimaging data shows promising accuracy	No detailed analysis of model interpretability or generalization across diverse datasets
[83]	AB, RF, DT, KNN, GNB, LR, SVM	Not used	Kaggle, UCI—Behavioral	1–18 + yrs/Mixed	AB best for toddlers, LDA for adolescents and adults	No explainability; data insufficiency
[84]	Rule-based ML	Not used	AGRE, Simons Simplex—Genetic + Behavioral	All ages/1804 samples	Easy-to-understand rules; high performance	Limited to adults; class imbalance
[85]	LR, NB, DT, SVM, ANN, RF	Not used	UCI—Behavioral + Genetic	All ages/1804 samples	DT and RF showed highest accuracy	No ensemble; explainability missing
[86]	SVM, Random Forest, Deep Learning	Not mentioned	Likely clinical/autism-related data	Various age groups	Empirical comparison of ML models for ASD diagnosis, highlighting accuracy and feature selection	Challenges with dataset variability and generalization across populations
[87]	SVM, RF, NB, LR, KNN	Not used	Autism Screening—Behavioral	Not specified/1054	Logistic Regression performed best	Imbalanced, small dataset
[88]	Modified GOA	Not used	UCI—Behavioral	1–18 + yrs/Mixed	Accuracy 100%	Explainability missing; convergence time is not fast enough
[89]	ad-DNN	SHAP	ABIDE—fMRI	5–40 yrs/Small set	Identified discriminative brain regions using SHAP	Small dataset; not clinically validated
[90]	Deep CNN	Grad-CAM	MRI Images—ABIDE	6–30 yrs/Small set	Identified ASD-related brain regions; high visual interpretability	No clinical validation; limited sample size
[91]	DNN, CNN, LR, KNN, SVM	SHAP	Public Dataset—Genetic + Behavioral	4–18 + yrs/1758	Top 7 features = 79% accuracy; outperformed others	Accuracy limited; lacks multiclass focus
[92]	XGBOOST	SHAP	Infant videos—Behavioral	9–18 mo/32 samples	AUCs of 0.938 and 0.914 from SHAP-based interpretation	Small, gender-imbalanced sample
[93]	CNN	LRP, PatternNet	Not specified—EEG data	15–18 yrs/88 samples	CNN outperformed humans in FER task	Small sample; only CNN tested
[94]	CNN, SVM	LIME, SHAP	Behavioral Video Data	3–8 yrs/100 samples	Improved visual understanding of social cues in ASD	Model overfitting; limited generalizability
[95]	LR, SVM, DT, RF, GB, NN	Not used	Not specified—Genetic	0–49 mo/240 samples	Ensemble of 742 models reached AUC-ROC ≥ 0.8	Small dataset; no psychometric analysis
[96]	CNN + LSTM	SHAP	EEG & Audio Fusion Dataset	4–10 yrs/80 samples	Fusion data led to better ASD prediction accuracy	Small, imbalanced dataset
[97]	XGBoost, RF, SVM	SHAP	Genetic Microarray Dataset	Infants/50 samples	SHAP highlighted top 5 ASD-related genes	Lacks replication on larger datasets
[98]	Custom DL + FL	LIME	Wearable Sensors—Behavior & Motion	Toddlers/60 samples	Preserved privacy with federated explainability	Complex deployment; data fragmentation
[99]	SVM, KNN, QDA, MLP	LIME + FL	ASD toddler dataset	Toddlers/Not specified	FL + XAI enhanced privacy and model accuracy	Small dataset; privacy/data heterogeneity
[100]	Custom Classifier	LRP	ABIDE—fMRI	15–18 yrs/1112 samples	High-order brain correlation mapping using relevance scores	Focused only on fMRI
[101]	Multiple (DT, XGB, AB, etc.)	Not used	Multi-modal—Socio/genetic/neuroimaging	12–36 mo/506 samples	AB classifier achieved 99.85% accuracy	Small dataset; lacks feature analysis

Recent research has successfully implemented XAI techniques in varying data types in the diagnosis of ASD, highlighting its potential in enhancing clinical decision-making. In the context of genomic data, the authors of [102] have used SHAP to explain support vector classifier (SVC) models, determining significant genes linked to ASD risk and thus presenting insights into genetic markers in tailoring treatment approaches. For neuroimaging information, the authors of [103] have employed fMRI scans using XGBoost and SHAP to identify brain areas with abnormal patterns of activity in ASD, e.g., posterior cingulate cortex, enabling clinicians to link brain activity to ASD symptoms. Within the domain of behavioral data, the authors of [104] have utilized LIME to explain machine learning models that examined social interaction patterns in ASD children and determined such important behaviors involved in ASD classification. In addition, in eye gaze analysis, the authors of [105] have used Grad-CAM to apply to gaze tracking data, allowing visual attention patterns in people with ASD to be identified, including decreased fixation on social stimuli, offering clinically relevant explanations that help to understand the causes of ASD behaviors. These case studies demonstrate how XAI techniques can render intricate data from various sources more understandable, aiding clinicians in making transparent, well-informed decisions for early ASD detection and intervention.

Comparison of existing XAI methods in ASD detection

Table 3 presents a comparative summary of prominent XAI methods applied in ASD diagnosis. It highlights the model compatibility, type of explanation offered, key strengths, and limitations of each technique, providing a concise reference for selecting appropriate XAI approaches in this domain.

Table 3.

Comparison of XAI techniques in the detection of ASD

XAI Method	Model Compatibility	Explanation Type	Strengths	Limitations	Use in ASD Research
LIME	Any black-box model	Local	Model-agnostic, easy to implement	Unstable explanations, lacks global insights	Used for interpreting behavioral models
SHAP	Tree-based, SVM, DNN, ensembles	Global & local	Consistent feature importance, handles interactions	Computationally expensive, complex for large datasets	Widely used for feature importance analysis
Grad-CAM	CNNs	Visual (image-based)	Highlights key image regions, useful in medical imaging	Only for CNNs, coarse resolution	Used in facial analysis studies
SmoothGrad	CNNs, DNNs	Noise-augmented gradients	Reduces visual noise in saliency maps, enhances clarity	Depends on gradient-based saliency, sensitive to input noise	Used in improving clarity of neuroimaging explanations
PatternNet	Neural networks	Signal extraction	Focuses on signal over noise, helps understand what is learned	Requires layer-wise computations, less intuitive	Applied in interpretability studies of neural patterns
LRP	Deep neural networks	Backpropagation-based	Provides detailed pixel/feature contributions, good for medical imaging	Model-specific implementation, can be complex to apply	Used in brain imaging and EEG-based ASD studies

Limitations and ethical concerns of XAI in clinical practice

Although XAI provides a potential advantage in healthcare, its actual deployment presents some challenges. One such concern is the trade-off between model performance and interpretability, as less complex, more interpretable models can be inferior to more complex black-box models. Additionally, most XAI techniques are still model-specific and not very generalizable, which restricts their suitability across various AI systems and disease conditions. The absence of quantitative measures to assess explanation quality also impedes uniform benchmarking and trust in such systems [106, 107].

Also, integration into clinical workflows necessitates user-friendly interfaces and collaboration with healthcare professionals to ensure explanations are actionable and meaningful. But clinical staff may not be trained in AI literacy, which can impede adoption and raise the risk of miscommunication. There is also the challenge of mapping XAI tools to diverse healthcare settings with different resource levels, data infrastructures, and regulatory requirements.

From an ethical standpoint, bias in training data may reinforce health disparities if not addressed appropriately. Privacy issues surrounding sensitive patient data are amplified in XAI systems, particularly where explanations reveal underlying patterns or attributes in the data. Unclear guidelines for guaranteeing explainability and fairness add more uncertainty and potential for abuse. In addition, there is the potential for overdependence on AI-produced insights, especially when healthcare professionals rely on algorithmic objectivity without a critical analysis of context. Issues regarding accountability in the event of misdiagnosis also arise where decisions are being made using AI, even when explanations are offered [108].

Addressing these challenges is necessary so that XAI not only remains technically sound, but also is ethically appropriate and clinically reliable. This will demand interdisciplinary engagement, regulatory architectures, and regular feedback loops among users to warrant secure and justifiable deployment into operational healthcare contexts.

Discussion

Even though research on ASD has advanced significantly, there are still a number of important issues that need to be resolved in order to provide reliable diagnostic tools. The interpretability and clinical acceptability of current ML models are limited by the inadequate integration of XAI techniques with genomic data analysis for early ASD detection. Furthermore, thorough studies of high-order functional correlations between different brain regions are lacking, which could help to clarify the mechanisms underlying ASD and guide the development of individualized treatment plans. Additionally, there is still a lack of widespread use of federated learning techniques for evaluating vast and heterogeneous datasets, which makes it difficult to achieve improved diagnostic scalability and accuracy while maintaining data privacy.

Conclusion

In conclusion, although it is both a technological and ethical necessity for XAI to be integrated into AI systems, its role in healthcare is especially crucial because of the high-stakes environment of clinical decision-making. This review has been focused on the significance of incorporating XAI into AI driven approaches for early detection of ASD—a neurodevelopmental disorder characterized by difficulty in social interaction and behavior. Through this analysis, several research gaps have been identified in existing studies, including the limited use of explainability tools in ASD detection, challenges related to small or imbalanced datasets, and the issue of clinical heterogeneity across studies. This article can serve as an exhaustive resource for researchers to develop interpretable, biomarker-based models for ASD, and can direct subsequent research into clinically meaningful AI applications in neurodevelopmental disorders. As part of future work, we aim to present a detailed study of various biomarkers that can be used in detection of ASD in children through transparent and clinically meaningful insights.

Abbreviations

AB

Ada boost classification algorithm

RF

Random forest classification algorithm

DT

Decision tree classification algorithm

KNN

K nearest neighbor’s classification algorithm

GNB

Gaussian naive Bayes classification algorithm

LR

Logistic regression classification algorithm

SVM

Support vector machine classification algorithm

LDA

Linear discriminant analysis

QDA

Quadratic discriminant analysis

NB

Naïve Bayes classification algorithm

CNN

Convolution neural network

ANN

Artificial neural network

GOA

Grasshopper optimization algorithm

DNN

Deep neural network

XGBOOST

Xtreme gradient boosting algorithm

CNN

Convolution neural network

RNN

Recurrent neural network

LSTM

Long short term memory algorithm

BiLSTM

Bidirectional long short-term memory algorithm

BERT

Bidirectional encoder representations from transformers

BERTweet

Bidirectional encoder representations from transformers for analyzing tweets

MLP

Multilayer perceptron

SHAP

SHapley additive exPlanations

LIME

Local interpretable model-agnostic explanations

LRP

Layer-wise relevance propagation

MLP

Multi layer perceptron classifier

GradCAM

Gradient-weighted class activation mapping

GGT

Graphical gene tree

SMOTE

Synthetic minority oversampling technique

GDPR

General data protection regulation

HIPAA

Health insurance portability and accountability act

PIPL

Personal information protection law

Author contributions

Renuka Agrawal: supervision, review and editing. Rucha Agrawal: conceptualization, investigation, writing-original draft, editing.

Funding

Open access funding provided by Symbiosis International (Deemed University).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Competing interests

The authors declare no competing interests.