Exploring the Intersection of Artificial Intelligence and Clinical Healthcare: A Multidisciplinary Review

Artificial intelligence (AI) plays a more and more important role in our everyday life due to the advantages that it brings when used, such as 24/7 availability, a very low percentage of errors, ability to provide real time insights, or performing a fast analysis. AI is increasingly being used in clinical medical and dental healthcare analyses, with valuable applications, which include disease diagnosis, risk assessment, treatment planning, and drug discovery. This paper presents a narrative literature review of AI use in healthcare from a multi-disciplinary perspective, specifically in the cardiology, allergology, endocrinology, and dental fields. The paper highlights data from recent research and development efforts in AI for healthcare, as well as challenges and limitations associated with AI implementation, such as data privacy and security considerations, along with ethical and legal concerns. The regulation of responsible design, development, and use of AI in healthcare is still in early stages due to the rapid evolution of the field. However, it is our duty to carefully consider the ethical implications of implementing AI and to respond appropriately. With the potential to reshape healthcare delivery and enhance patient outcomes, AI systems continue to reveal their capabilities.


What Is AI?
Artificial intelligence (AI) is the simulation of human intelligence processes by machines or computer systems. Natural language processing, speech recognition, expert systems, or machine vision are common applications of AI. In simple words, the AI field combines computer science and good-quality, vetted, datasets for the purpose of solving a given problem. It also encompasses sub-fields of machine learning and deep learning, which are frequently mentioned in conjunction with artificial intelligence [1]. These disciplines are comprised of AI algorithms that seek to create expert systems, which make predictions or classifications based on input data. Over the years, artificial intelligence has gone through many cycles of hype, but even to skeptics, the ChatGPT release of OpenAI seems to mark a turning point [2]. The last time generative AI loomed this large, the breakthroughs were in computer vision [3], but now the leap forward is in natural language

Artificial Intelligence Algorithms in Medicine
AI algorithms appear to be magical, but they are simply mathematical functions that describe input data and map them to outputs. The input data in healthcare includes structured elements such as diagnostic codes, vital sign fields, and demographic fields, as well as unstructured data such as text and radiologic images. Input data should be selected based on features relevant to the desired prediction task. A variety of models must be applied to test for the best performance since inputs can be highly variable with no clear interrelationships between them [28].
The four fundamental tasks listed under the term of competence, i.e., diagnosing, estimating a prognosis, identifying causes of diseases, and selecting treatments, must be gained from patients prior to a diagnosis. It was not until the twentieth century that data were collected beyond medical histories and physical examinations conducted in the privacy of medical offices. Following this, data collection underwent an industrial revolution, characterized by machine use and division of labor [29]. As medical decisions become more complex due to the proliferation of data sources and the need to involve multiple specialists, physicians have shifted from making individual decisions in the privacy of their offices to making collective decisions in multidisciplinary meetings.
After the first step of collecting data from their patients, the second step for physicians is to use their clinical reasoning to make medical decisions. From an informatics point of view, clinical reasoning is data processing; data can be processed by algorithms, and algorithms are currently able to deliver a diagnostic probability, a prognostic estimation, or the selection of a treatment. As data collection and data processing/AI have progressed, the number of actors involved in patient care has multiplied, and these actors are no longer limited to humans Diagnostics 2023, 13, 1995 5 of 37 but also software that makes medical decisions. In order to maintain their "competence" dimension, physicians must retain control over these new technologies [30].
AI tools fit along a continuum between fully human and fully computer-driven on the concept of a human-computer collaboration spectrum. This view emphasizes levels of analytic complexity that provide a framework for clarifying forms of machine learning. In contrast, a deep learning model can automatically classify inputs without much human intervention, unlike a classic statistical model. Figure 1 synthesizes the main methods and algorithms of machine learning used in medicine: (1) Supervised learning: These models are trained on labeled data to learn patterns and make predictions. They are widely used in medical image analyses, such as identifying cancer cells in pathology images or detecting lung nodules in CT scans [7]. Some examples of supervised learning models used in medicine include convolutional neural networks (CNNs), deep neural networks (DNNs), and random forests. (2) Unsupervised learning models: These models are used to identify patterns and relationships in unlabeled data. They are used in medical data clustering, anomaly detection, and feature extraction. Some examples of unsupervised learning models used in medicine include k-means clustering, a principal component analysis (PCA), and autoencoders [31]. (3) Reinforcement learning models: These models are used to learn from trial-and-error interactions with an environment. They can be used in medical decision making, such as personalized treatment planning and drug dosage optimization [32]. Examples of such models include Q-learning, policy gradient methods, and actor-critic models. (4) Hybrid models: These models combine multiple types of AI models to leverage their strengths and overcome their weaknesses. For example, a hybrid model could use a CNN to identify features in medical images followed by an unsupervised learning algorithm to cluster the features and identify subtypes of cancer [33].
Diagnostics 2023, 13, x FOR PEER REVIEW 6 of 37 The main AI systems used in medicine are synthesized in Table 1. Table 1. Main AI systems used in medicine.

Abbreviation Function
Artificial Neural Network ANN It is trained by processing examples, each of which contains a known "input" and "result," forming probability-weighted associations between the two, which are stored within the data structure of the net itself [34].
Backpropagation Neural Network -Backpropagation is a process involved in training a neural network. It involves taking the error rate of a forward propagation and feeding this loss backward through the neural network layers to fine-tune the weights. Backpropagation is the essence of neural net training The main AI systems used in medicine are synthesized in Table 1.

System Abbreviation Function
Artificial Neural Network ANN It is trained by processing examples, each of which contains a known "input" and "result," forming probability-weighted associations between the two, which are stored within the data structure of the net itself [34].

Backpropagation Neural Network -
Backpropagation is a process involved in training a neural network. It involves taking the error rate of a forward propagation and feeding this loss backward through the neural network layers to fine-tune the weights. Backpropagation is the essence of neural net training [35].
Bayesian Inference -It allows for an algorithm to make predictions based on prior beliefs. In Bayesian inference, the posterior distribution of predictors (derived from observed data) is updated based on new evidence [36].

Causal Associational Network CASNET
This model consists of three main components: observations of a patient, pathophysiological states, and disease classifications. As observations are recorded, they are associated with the appropriate states [37].

Convolutional Neural Network CNN
A network architecture for deep learning that learns directly from data. CNNs are particularly useful for finding patterns in images to recognize objects, classes, and categories. They can also be quite effective for classifying audio, time-series, and signal data [38].

Deep Neural Network DNN
An ANN with multiple layers between the input and output layers. There are different types of neural networks but they always consist of the same components: neurons, synapses, weights, biases, and functions [39].

Light Gradient Boosting Machine LightGBM
LightGBM is a gradient-boosting ensemble method that is based on decision trees. As with other decision tree-based methods, LightGBM can be used for both classification and regression. LightGBM is optimized for a high performance with distributed systems [40].

Multilayer Perceptron MLP
A feedforward artificial neural network that generates a set of outputs from a set of inputs. An MLP is characterized by several layers of input nodes connected as a directed graph between the input and output layers. MLP uses backpropagation for training the network [41].

Natural Language Processing NLP
It enables machines to understand the human language. Its goal is to build systems that can make sense of text and automatically perform tasks such as translation, a spell check, or topic classification [18].

Optimal Channel Networks OCNet
Oriented spanning trees that reproduce all scaling features characteristic of real, natural river networks. As such, they can be used in a variety of numerical and laboratory experiments in the fields of hydrology, ecology, and epidemiology [42].

AI in Cardiology
AI found its applications in cardiology in several domains, such as imaging, electrophysiology, and heart failure prediction, as well as preventive and interventional cardiology. AI models can be used for an electrocardiogram (ECG) analysis. Deep learning models have been shown to accurately identify cardiac arrhythmias, including atrial fibrillation (AF), and ventricular tachycardia, from ECG recordings [50]. A growing number of people are likely to develop AF within the next three decades [51], making early diagnoses and management critical in primary care. By screening for undetected AF in primary care populations, ML models may enable early anticoagulation and reduce the subsequent disease burden. For the diagnosis of AF, ML models are trained on clinical data from electronic patient records (EPRs) or ECGs [52]. By using the former approach, clinicians can assist patients with screening based on age, previous cardiovascular disease, and body mass index. A ML model based on an ECG waveform analysis has demonstrated high accuracy. On the basis of 18,000 ECG signals, a deep learning system can diagnose atrial fibrillation with an accuracy of 98.27% [53]. Studies have demonstrated that many general practitioners are incapable of accurately detecting and diagnosing AF based on ECGs [9]. ECGs may be useful in identifying high-risk patients, subsequently resulting in a combined approach pathway. Any non-trivial traces can then be identified or flagged for specialist intervention.
ML models have also been used to automatically identify and quantify cardiac function parameters, such as the ejection fraction, without the need for manual measurements [16]. AI can also assist in the prediction of the cardiovascular disease (CVD) risk. A study by Weiss et al. [54] demonstrated that a deep learning model can predict the 10-year risk of cardiovascular disease more accurately than traditional risk calculators, based on a single chest X-ray. Additionally, AI can help identify patients who may benefit from preventive interventions or close monitoring, potentially reducing the risk of adverse cardiovascular events. The main outcomes of AI use in cardiology are presented in Table 2. AI has also been applied in the prediction of patient outcomes. AI algorithms can analyze specific parameters to predict the likelihood of a patient developing certain complications, such as heart failure or a stroke. Hamatani et al. [55] used a ML model based on a random forest algorithm to assess the heart failure hospitalization in patients with atrial fibrillation. The proposed model exerted a higher prediction performance than the Framingham risk model [55]. In a systematic review, Kee et al. [56] observed that a neural network was able to predict the risk of CVD in type 2 diabetes patients, with 76.6% precision and 88.06% sensitivity [56].    A self-taught ML model was found to be better at predicting the risk of death in CVD patients than other models designed by cardiovascular experts [73]. Samad et al. [70] used echocardiographic and clinical parameters on a supervised learning model in order to obtain the survival prediction, compared to other risk scores and logistic regression models.
Personalized treatment is another area where AI can be useful in cardiology. AI algorithms can analyze patient data to identify the most effective treatment options for individual patients based on their unique characteristics. Chi et al. [74] used a machine learning personalized statin treatment plan to assess the available statin plans and to identify the optimal treatment plan in order to prevent or minimize statin patient discontinuation.

AI Implications in Immunology, Allergology, and COVID-19
The potential clinical applications of AI in allergies and immunology have a wide range, from a common disease diagnosis (food allergy, asthma, and drug allergy) to diseases with a delayed diagnosis, which fail to be very obvious from the beginning to the general practitioners and pediatricians, thereby endangering newborns' lives, including the inborn errors of immunity. Other potential clinical applications include the assessment and prediction of adverse reactions to drugs and vaccines-to the pandemic pathology and the post-vaccination immune response to coronavirus disease 2019 (COVID-19) and non-COVID-19, and to the multidimensional data reduction in the electronic field, the health records, and the immunological datasets.
One area where AI has been applied in allergology is in the diagnosis of allergic diseases. AI algorithms can analyze patient data, such as medical history, allergy testing results, and environmental exposure data, to identify patterns and associations that may be indicative of allergic diseases. Yang et al. developed an ensemble neural network chain model with pre-training on rhinitis multi-label classification. Malizia et al. [75,76] established a machine learning model that, based on nasal cytology and skin prick test results, could identify allergic rhinitis phenotypes in children. Nevertheless, the authors acknowledge that cytologic endotypes over time may limit the efficiency of such a model [75]. Bhardwaj et al. [77] successfully trained and tested six ML models to classify allergic and non-allergic asthma.
AI methods have also been used in the prediction of allergic diseases and complications. Research conducted by van Breugel focused on a multi-omics model, which could accurately perform a methylation-based allergy diagnosis [78]. Therefore, ML models are able to go beyond a simple analysis of one domain and to integrate multi-omics layers [79]. AI algorithms can analyze patient data to predict the likelihood of a patient developing severe allergic reactions, such as anaphylaxis [80].
Another important aspect that could gain a major benefit from AI and ML is the discovery of drug allergies by establishing a risk profile for patients at risk of developing a drug allergy. The most common example is the beta-lactam amoxicillin and clavulanic acid combination, responsible in the last decade for late and immediate hypersensitivity reactions [81].
AI has been applied in the development of personalized treatment plans. AI algorithms can identify, based on patient data, the most effective individual treatment options [82].
ML endeavors to accomplish precision medicine in allergology by characterizing allergic endotypes, exploring relationships in allergic multimorbidity, contextualizing the impact of an exposome, and intervening in biological processes to enhance health and treat individual diseases. Exposure represents a critical factor in the allergic disease physiopathology, a high complexity factor mainly due to the possibility of multiple exposures that can occur simultaneously. The concept of an "exposome" has been introduced, a term that encloses "all exposures from conception onwards" [83]-a complex puzzle that can be put together by ML algorithms ( Figure 2). Nevertheless, the beneficial role of ML in the exposome investigation is closely related to the quality of analyzed data. Contemporary ML techniques employ embeddings to transform high-dimensional feature spaces into efficient representations. These approaches often leverage modern DL methods such as convolutional neural networks [84]. While these approaches demonstrate high prediction accuracy, it is essential to acknowledge that the patterns identified by these methods may be merely correlative, lacking direct associations with the underlying molecular mechanism. Nevertheless, they remain valuable as biomarkers in clinical assessments.
A convolutional neural network was used to accurately identify and count airborne pollen, to distinguish between the low-allergenic Urtica species and severely allergenic Parietaria species [15]. The authors observed that Urticaceae pollen grains could be distinguished with >98% accuracy. Moreover, the model could distinguish genera on before unseen Urticaceae pollen collected from aerobiological samples [15]. Olsson et al. [85] trained CNN models on 122,000 pollen grains, obtaining an accuracy of 98% for 83 species. Nevertheless, the accuracy dropped to 41% when individual reference samples from different flowers were kept separate [85]. Samonte et al. [86] developed a web-based application for food recommendation specialized in allergy information. In this application, restaurants would upload their menu and the individuals could make their choices based on potential known allergies. A selection of ML models' outcomes in allergology is presented in Table 3. Table 3. Outcomes of ML models in allergology. Contemporary ML techniques employ embeddings to transform high-dimensional feature spaces into efficient representations. These approaches often leverage modern DL methods such as convolutional neural networks [84]. While these approaches demonstrate high prediction accuracy, it is essential to acknowledge that the patterns identified by these methods may be merely correlative, lacking direct associations with the underlying molecular mechanism. Nevertheless, they remain valuable as biomarkers in clinical assessments.
A convolutional neural network was used to accurately identify and count airborne pollen, to distinguish between the low-allergenic Urtica species and severely allergenic Parietaria species [15]. The authors observed that Urticaceae pollen grains could be distinguished with >98% accuracy. Moreover, the model could distinguish genera on before unseen Urticaceae pollen collected from aerobiological samples [15]. Olsson et al. [85] trained CNN models on 122,000 pollen grains, obtaining an accuracy of 98% for 83 species. Nevertheless, the accuracy dropped to 41% when individual reference samples from different flowers were kept separate [85]. Samonte et al. [86] developed a web-based application for food recommendation specialized in allergy information. In this application, restaurants would upload their menu and the individuals could make their choices based on potential known allergies. A selection of ML models' outcomes in allergology is presented in Table 3. ML learning frameworks have been developed specifically for allergy diagnoses, aiming to support junior clinicians and specialists in their decision-making tasks [94]. The main objective was to assist the management of complex cases, with multiple allergies, rather than focusing on easily diagnosable primary allergies. The framework includes a data cleaning module and utilizes modified sampling techniques in the data sampling module to improve the quality of intradermal test data. These processing steps significantly enhance the performance of the learning algorithms. Moreover, the adoption of a crossvalidation approach ensures that the learning algorithms avoid overfitting the training data. Notably, ensemble classification approaches demonstrate a superior performance compared to traditional methods. The random forest classifier, employing constant strategy sampling, demonstrated superior sensitivity compared to all other cases [94].
To further improve the efficiency of the allergy diagnosis support system, metaheuristic data-processing techniques can be employed. In addition to data cleaning and sampling, incorporating data transformation methods, such as feature selection, can be beneficial. Including prognosis details, treatment outcomes, and patient feedback will enhance the relevance of the system. ML algorithms presented high accuracy and efficiency in identification of systemic lupus erythematosus (SLE) and neuropsychiatric systemic lupus erythematosus [95,96], as well as distinguishing patients with SLE and other major chronic autoimmune diseases, such as rheumatoid arthritis and multiple sclerosis, in the early stages [97]. Ali et al. [8] used a transcriptomic fragmentation model for biomarker detection in multiple sclerosis and rheumatoid arthritis, with a 96.45% accuracy. Li et al. [98] proposed combined proteomics and single-cell RNA sequencing to determine biomarker combinations for the diagnosis and activity monitoring in SLE patients; their model could efficiently assess disease exacerbation [98].
Martin-Gutierrez et al. [99] employed ML models to identify distinct immunologic signatures in subjects with primary Sjögren's syndrome and SLE. The proposed model identified two distinct immune cell profiles, which could provide further directions in targeted therapy [99]. Therefore, AI can also be used to discover new treatments and predict drug efficiency for immune diseases by analyzing large amounts of genomic and proteomic data. ML algorithms predicted the efficiency of the etanercept in juvenile idiopathic arthritis using electronical medical records data, with a 75% sensitivity and 66.67% specificity [100].
Based on deep learning algorithms, Zeng et al. [101] developed deepDTnet, a model for target identification and drug repurposing, enclosing 15 types of cellular, phenotypic, genomic, and chemical profiles. Their proposed model showed a 96.3% accuracy in identifying novel molecular targets for known drugs [101]. Madhukar et al. [102] promoted BANDIT, a ML model that integrates multiple data types to identify connections between different drug types and classes and to predict drug binding targets.
AI has shown great potential in the field of vaccine development, as it can help to accelerate the identification of potential vaccine targets and the development of new vaccine candidates. Bukhari et al. [103] proposed a decision tree model for the prediction of novel and immunodominant Zika virus T-cell epitopes. The model showed a mean accuracy of 97.86%, with high possibilities in the development of Zika vaccines that target the predicted T-cell epitopes [103]. Arterolane and lucanthone were identified, based on a Bayesian ML model, as potential Ebola virus inhibitory agents [104].
AI-based models were also used for the COVID-19 vaccine development. Neural network-driven systems were used to discover T-cell epitopes for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) [105]. The Long Short-Term Memory network was used to predict epitopes for Spike [106]. Pre-trained models were also used to predict molecular reactions in carbohydrate chemistry [107] and protein interaction [108].
Medical advances and high-tech developments, including AI, have led to significant advances in treating COVID-19. As a consequence of the inability to accurately and efficiently evaluate pulmonary lung CT data during the fight against COVID-19, Zhang et al. [109] developed a new system to analyze CT data of patients using deep learning and concluded that the right lower lobe of the lungs is the most common location for COVID-19 pneumonia. Additionally, Mohanty et al. [110] performed a quick intelligent screening for potential drugs to treat COVID-19 with a drug-repositioning method; this group was able to identify potential drugs using a combination of artificial intelligence and pharmacology, demonstrating the usefulness of this method to COVID-19 drug design and research. Moreover, other scholars have developed a platform based on AI learning and prediction models to identify the drugs on the market that may be useful for treating COVID-19; as a result, they found more than 80 drugs with considerable potential [111]. Stebbing et al. [112] analyzed existing anti-cytokine therapies, such as baricitinib, to explore new treatment options for COVID-19. Table 4 summarizes the developments of AI models in COVID-19 diagnoses, treatment, and prevention.  [123] Integrated bioinformatics pipeline that merges the prediction power of different software (in silico pipeline).
Predict the cross-reactivity of pre-existing vaccination interventions against SARS-CoV-2.
Three epitope-based subunit vaccines were designated. Only one was reported as the best vaccine.
A multi-epitopic vaccine candidate targeting the non-mutational immunogenic regions in envelope protein and surface glycoprotein of SARS-CoV-2.

AI in Endocrinology
AI models have been analyzed in the diagnosis and treatment of multiple endocrinological conditions and pathologies, such as diabetes, thyroid disorders, reproductive impairments, or hormonal cancers.
One example of the use of AI in diabetes management is the development of closedloop systems, also known as artificial pancreas systems. These systems use a combination of continuous glucose monitoring and insulin pumps to automatically regulate blood glucose levels. Neural network models had the most stable performance in such systems, being able to recover dynamics in short time intervals [127].
AI can also assist in the prediction of hypoglycemia, a common complication of diabetes. Continuous glucose monitoring data and clinical parameters are used in ML models to improve hypoglycemia prediction [128]. Ma et al. [129] introduced the MMTOP (multiple models for missing values at time of prediction) algorithm to predict the patient risk for severe hypoglycemia in the presence of incomplete data, with a cross-validated concordance index of 0.77 ± 0.03. Faruqui et al. [130] used deep learning algorithms to predict glucose levels in type 2 diabetes patients based on their diet, weight, glucose level from the day before, and physical activity.
Deep neural networks were able to predict gestational diabetes in early pregnancy, based on 73 variables, such as body mass index, 3,3,5 -triiodothyronine, or total thyroxin [131]. An unsupervised ML model was used to accurately classify four stable metabolic different obesity clusters: metabolic healthy obesity, hypermetabolic hyperuricemic obesity, hypermetabolic hyperinsulinemic obesity, and hypometabolic obesity [132]. Rein et al. [133] assessed the effect of a personalized postprandial-targeting diet (PPT) on glycemic control and metabolic health. The 6-month PPT intervention exerted significant improvement on glycated hemoglobin, fasting glucose, and triglycerides [133].
CNNs were used in thyroid pathology diagnoses. Yang et al. [134] proposed a deep learning framework trained on 508 ultrasound images to diagnose thyroid nodules. Their model showed an average accuracy of 98.4%. Islam et al. [135] compared 11 ML algorithms for thyroid risk prediction; the neural network classifier generated the highest accuracy over other ML techniques.
Reproductive health is a critical aspect of overall health and well-being, affecting individuals of all ages and genders. Hormonal imbalances and reproductive disorders can lead to infertility, pregnancy complications, and other health issues. In recent years, AI has emerged as a powerful tool for providing insights that may not be visible to human analyses. Polycystic ovary syndrome (PCOS) is a common hormonal disorder that affects up to 10% of women of a reproductive age [136]. It is characterized by irregular menstrual cycles, high levels of male hormones, and ovarian cysts [137]. The diagnosis of PCOS is currently based on clinical symptoms and a laboratory test, which can be subjective and lead to a misdiagnosis [138]. ML algorithms can assist in the diagnosis and management of PCOS. Suha and Islam [139] trained a CNN model on 594 ovary ultrasound images for cyst detection and PCOS diagnoses, with an accuracy of 99.89%. Zigarelli et al. [140] developed a self-diagnostic prediction model for PCOS, based on different variables, such as hirsutism, acne, an irregular menstrual cycle, the length of the menstrual cycle, and weight gain. Their model predicted a correct diagnosis with an accuracy ranging from 81% to 82.5%. Even if such self-diagnosis models can be useful in particular cases, which may include a lack of access to medical care or pandemic confinement, we consider that they should be taken "with a grain of salt", as they cannot replace a professional diagnosis.
AI can also assist in the diagnosis and management of infertility by providing personalized recommendations. Ding et al. [141] compared seven ML methods in order to assess the ovarian reserve. The most accurate evaluation was provided by a light gradient boosting machine (LightGBM), which exerted the highest accuracy in the quantification of the ovarian reserve, especially in the 20-35 years age group [141]. The basal body temperature and heart rate were used to train ML algorithms, in order to predict the fertile window (a 72.51% accuracy) and menses (75.90% accuracy) [142].
A deep CNN was trained using single timepoint images of embryos, with an accuracy of 90% in choosing the highest quality embryo for in vitro fertilization (IVF); the trained CNN was also capable of predicting the potential of embryo implantation [143]. Louis et al. [144] analyzed different ML models (decision tree, random forest, and gradient booster) for IVF embryo selection. Their result revealed a lower peak accuracy of 65% [144].

AI in Dentistry
In recent years, AI has gained significant attention in the field of dentistry, with many researchers exploring its potential applications in diagnoses, treatment planning, and dental imaging analyses.
One of the most promising applications of AI in dentistry is in the area of dental imaging analyses; the use of AI algorithms has the potential to improve the accuracy and speed of an image analysis, with the identification of early signs of carious lesions, periapical lesions, or periodontal destruction. Furthermore, AI models can be trained to detect subtle changes in images over time, which may be indicative of the disease progression. Ameli et al. [145] used ordinal logistic regression and artificial neural networks to determine predictive relationships between the extracted patient chart data topics and oral health-related contributors; the authors observed that the risk for carious lesions, occlusal risk, biomechanical risk, gingival recession, periodontitis, and gingivitis were highly predictable using the extracted radiographic and treatment planning topics and chart information [145].
Carious lesions are usually detected by a clinical examination and X-ray visual analysis, highly depending on the experience of the specialists. Numerous studies focused on the AI models' role in the early detection of carious lesions on dental X-rays (Table 5). Kühnisch et al. [146] proposed a CNN algorithm for carious lesion diagnoses on intraoral X-rays. Another study [147] compared the cost-effectiveness of AI for the detection of proximal caries with the diagnosis of dentists; the authors observed that the AI system was more effective and less expensive. Furthermore, AI algorithms can detect and analyze subtle changes in the periapical area, root canal anatomy, and bone structure [148].
A systematic review conducted by Mohammad-Rahimi et al. in 2022 assessed the DL capacity in various studies to detect carious lesions [149]. The authors observed different accuracies, mainly depending on the type of dataset, but with relatively high values: 71% to 96% on intra-oral photographs, 82% to 99.2% on periapical radiographs, 87.6% to 95.4% on bitewing radiographs, 68.0% to 78.0% on near-infrared transillumination images, 88.7% to 95.2% on optical coherence tomography images, and 86.1% to 96.1% on panoramic radiographs.  AI also found its way in periodontal diagnoses and prognoses [159][160][161]. AI models were used in order to detect the periodontal bone loss [159], periodontally compromised teeth [162], and even periodontal condition examination [163]. Troiano et al. [164] analyzed different AI models' efficiency in assessing overall molar loss in patients after active periodontal treatment, with favorable results. A synthesis of the main outcomes in periodontics is presented in Table 6. There is also great potential in AI for type recognition, success recognition, prediction, design, and optimization in dental implantology, as demonstrated by Revilla-Leon et al. [172]. AI systems can assist dentists and oral surgeons in planning the placement of dental implants by analyzing CBCT (cone-beam computed tomography) scans and identifying the optimal location, angulation, and size of implants, with a reduced risk of errors and complications [173]. AI can assist in the precise placement of dental implants during surgery by providing real-time guidance and feedback to clinicians (Table 7). AI can analyze CBCT scans and intraoperative data to help clinicians navigate the surgical site and ensure that the implants are placed in the optimal location and angulation [174]. AI can help dental professionals to design and create more accurate and personalized dental prosthetics for patients, by analyzing CBCT scans and digital impressions to create virtual 3D models. AI algorithms can also help to optimize the shape, size, and color of the restoration, ensuring a precise fit and a natural-looking appearance [179].
In orthodontics, AI systems have been applied in treatment planning and prediction of treatment outcomes, such as simulating changes in the appearance of pre-and post-treatment facial photographs [180] (Table 8). AI algorithms have been used in assessing the impact of orthodontic treatment, skeletal patterns, and anatomical landmarks in lateral cephalograms [181]. Other applications involved the diagnosis of the need for orthodontic treatment, tooth extraction determination in orthodontic treatments, or skeletal classification [182][183][184][185][186][187][188][189]. In oral and maxillofacial pathology, AI has been mainly researched for tumor and cancer detection based on radiographic, microscopic, and ultrasonographic images (Table 9). CNN models proved their accuracy end efficiency in detecting oral cancers [190]. Hung et al. [191] reviewed machine learning algorithms to predict oral cancer survival and factors affecting it; the authors concluded that cancer survival prediction and medical decision making were possible with the help of AI systems. Table 9. AI outcomes in oral and maxillofacial surgery.

Target AI Model Sample Results Study
Lower-third-molar treatment-planning decisions Neural networks Data from 119 patients Sensitivity of 0.78, which was slightly lower than the oral surgeon's (0.88), but the difference was not significant, and a specificity of 0.98, which was lower than the oral surgeon's (0.99) (p = NS). Brickley and Shepherd (1996) [192] To predict postoperative facial swelling following impacted mandibular third molar extraction Poedjiastoeti and Suebnukarn (2018) [194] [198]. The model was trained on 240 radiographs (120 with intact dental roots and 120 with vertically fractured roots), as well as on conebeam computed tomographies (CBCTs). The maximum accuracy, sensitivity, and specificity values in the three groups were 70.00, 97.78, and 67.7%, respectively, for radiographic images. When using CBCT images, the values were 96.6, 93.3, and 100%, respectively. Table 10. AI outcomes in endodontics.

AI Model Sample Results Study
Locating the minor apical foramen ANN 50 teeth To enhance the accuracy of working length measurement using radiography, artificial neural networks can serve as a second opinion to find the apical foramen on radiographs.
Saghiri et al., Therefore, various types of AI models are currently employed in the field of dentistry. Neural networks, including CNNs and ANNs, were among the earliest AI algorithms used. CNNs are primarily utilized for analyzing dental images. However, it is essential to implement more robust, reproducible, and standardized processes in the future, to ensure the usefulness, security, and widespread applicability of these models.

AI Advantages in Medicine
Healthcare programs and procedures can benefit from the AI systems; the main advantages of using AI in medicine are presented in Figure 3. AI algorithms can process vast amounts of patient data and help physicians make more accurate and timely diagnoses. Therefore, they can reduce the risk of a misdiagnosis and improve patient outcomes [203], as well as reduce the initial process time up to 70% [204].

AI Advantages in Medicine
Healthcare programs and procedures can benefit from the AI systems; the main advantages of using AI in medicine are presented in Figure 3. AI algorithms can process vast amounts of patient data and help physicians make more accurate and timely diagnoses. Therefore, they can reduce the risk of a misdiagnosis and improve patient outcomes [203], as well as reduce the initial process time up to 70% [204]. The 21st century belongs to personalized medicine, in which AI can play an important part. AI can help doctors tailor treatment plans to individual patients by analyzing patient data, medical records, and other relevant information.
AI algorithms can help healthcare providers identify patients who are at risk of developing complications or adverse reactions to treatment, allowing for early intervention and improved outcomes. Different AI models can automate many routine tasks, freeing up physicians and other healthcare professionals to focus on more complex cases and improving overall efficiency [205]. Continuous monitoring plays a crucial role in preventing potentially dangerous situations, allowing for the fine-tuning of ongoing treatments. This proactive approach enables a reduction of up to 40% in the total duration from the onset of illness to complete recovery. Moreover, AI facilitates the planning of more effective treatments while also accelerating the research and development of new medicines [206,207].
The ability to make early predictions holds tremendous potential in enhancing medical care, enabling healthcare providers to deliver more effective treatments and interventions. By leveraging intelligent phone-based prediction systems, patients gain the convenience of assessing their health condition without the need for in-person visits to the hospital. These systems use advanced algorithms and data analysis techniques to analyze symptoms, medical history, and other relevant factors, providing individuals with valuable insights into their current data [208].
Furthermore, AI can aid in identifying the root cause of various diseases. By analyzing a wide range of data, including genetic information, lifestyle factors, and environmental influences, these systems can uncover factors contributing to the development and progression of diseases [209].
Even though AI programs can be expensive, a global and perspective image of using AI mechanisms can generate cost savings in the end. This can be explained by the improved efficiency, a reduced risk of medical error, and a minimization of the need for expensive procedures and tests [210]. The 21st century belongs to personalized medicine, in which AI can play an important part. AI can help doctors tailor treatment plans to individual patients by analyzing patient data, medical records, and other relevant information.
AI algorithms can help healthcare providers identify patients who are at risk of developing complications or adverse reactions to treatment, allowing for early intervention and improved outcomes. Different AI models can automate many routine tasks, freeing up physicians and other healthcare professionals to focus on more complex cases and improving overall efficiency [205]. Continuous monitoring plays a crucial role in preventing potentially dangerous situations, allowing for the fine-tuning of ongoing treatments. This proactive approach enables a reduction of up to 40% in the total duration from the onset of illness to complete recovery. Moreover, AI facilitates the planning of more effective treatments while also accelerating the research and development of new medicines [206,207].
The ability to make early predictions holds tremendous potential in enhancing medical care, enabling healthcare providers to deliver more effective treatments and interventions. By leveraging intelligent phone-based prediction systems, patients gain the convenience of assessing their health condition without the need for in-person visits to the hospital. These systems use advanced algorithms and data analysis techniques to analyze symptoms, medical history, and other relevant factors, providing individuals with valuable insights into their current data [208].
Furthermore, AI can aid in identifying the root cause of various diseases. By analyzing a wide range of data, including genetic information, lifestyle factors, and environmental influences, these systems can uncover factors contributing to the development and progression of diseases [209].
Even though AI programs can be expensive, a global and perspective image of using AI mechanisms can generate cost savings in the end. This can be explained by the improved efficiency, a reduced risk of medical error, and a minimization of the need for expensive procedures and tests [210].
Moreover, AI systems can improve resource allocation. They can help healthcare providers identify areas where resources are needed most, such as high-risk patient populations or under-resourced communities [211].
Another potential advantage involves the accelerated drug discovery in which AI can be beneficial. AI algorithms are able to analyze large amounts of data to identify potential drug candidates and speed up the drug discovery process [32].

AI Disadvantages and Limitations in Medicine
Although AI has the potential to revolutionize healthcare and improve patient outcomes, there are also several disadvantages and limitations to its use in medicine (as synthesized in Figure 4). One of the major concerns with AI in healthcare is the lack of trust and transparency in the decision-making process [212]. Both healthcare providers and patients may be hesitant to rely on AI algorithms for a critical decision without a clear understanding of how the algorithm reached its conclusion. The level of trust that individuals have in AI is influenced by a diverse array of human characteristics. Factors such as education, user preferences, life experiences, and attitudes toward automation can all play a role in shaping trust [213]. People with a higher level of education or those who had positive experiences with AI technologies may be more inclined to trust AI systems. Moreover, AI systems can improve resource allocation. They can help healthcare providers identify areas where resources are needed most, such as high-risk patient populations or under-resourced communities [211].
Another potential advantage involves the accelerated drug discovery in which AI can be beneficial. AI algorithms are able to analyze large amounts of data to identify potential drug candidates and speed up the drug discovery process [32].

AI Disadvantages and Limitations in Medicine
Although AI has the potential to revolutionize healthcare and improve patient outcomes, there are also several disadvantages and limitations to its use in medicine (as synthesized in Figure 4). One of the major concerns with AI in healthcare is the lack of trust and transparency in the decision-making process [212]. Both healthcare providers and patients may be hesitant to rely on AI algorithms for a critical decision without a clear understanding of how the algorithm reached its conclusion. The level of trust that individuals have in AI is influenced by a diverse array of human characteristics. Factors such as education, user preferences, life experiences, and attitudes toward automation can all play a role in shaping trust [213]. People with a higher level of education or those who had positive experiences with AI technologies may be more inclined to trust AI systems. Additionally, trust in AI is also influenced by the characteristics and attributes of the AI systems themselves. The degree of control that users have over the AI systems can impact trust levels. Users are more likely to trust AI systems that allow them to understand and influence the decision-making process. Transparent AI systems that provide clear explanations of their actions and reasoning can also enhance trust. On the other hand, highly complex AI systems that are difficult to comprehend may lower trust levels [214].
If users perceive AI systems to be prone to errors or potential harm, their trust may also be diminished. Ensuring the security and privacy of personal data handled by AI systems is essential for building trust.
Educating users about the capabilities and limitations of AI systems can favor trust levels. Developing systems with user-centric designs that prioritize transparency, explainability, and control can also foster trust. Additionally, addressing the ethical and regulatory concerns surrounding AI and implementing robust measures for data protection can enhance trust in AI technologies.
Another important limitation of AI in medicine is the need for large amounts of highquality data to train AI algorithms [215]. The data must be carefully collected and curated to ensure that it is representative and unbiased. However, there may be challenges in Additionally, trust in AI is also influenced by the characteristics and attributes of the AI systems themselves. The degree of control that users have over the AI systems can impact trust levels. Users are more likely to trust AI systems that allow them to understand and influence the decision-making process. Transparent AI systems that provide clear explanations of their actions and reasoning can also enhance trust. On the other hand, highly complex AI systems that are difficult to comprehend may lower trust levels [214].
If users perceive AI systems to be prone to errors or potential harm, their trust may also be diminished. Ensuring the security and privacy of personal data handled by AI systems is essential for building trust.
Educating users about the capabilities and limitations of AI systems can favor trust levels. Developing systems with user-centric designs that prioritize transparency, explainability, and control can also foster trust. Additionally, addressing the ethical and regulatory concerns surrounding AI and implementing robust measures for data protection can enhance trust in AI technologies.
Another important limitation of AI in medicine is the need for large amounts of highquality data to train AI algorithms [215]. The data must be carefully collected and curated to ensure that it is representative and unbiased. However, there may be challenges in collecting and sharing data across different healthcare systems due to issues of privacy, data ownership, and regulatory compliance. In addition, AI systems can be biased towards certain groups such as those with more available data [216]. This can lead to inaccuracies in diagnoses and treatment plans for underrepresented populations. Moreover, AI models can misinterpret data, leading to incorrect diagnoses or treatment plans. This is especially true when the data is noisy or incomplete, which is often the case in healthcare [5].
The development and implementation of AI systems can generate significant costs. This aspect can limit the access to these technologies, in particular in low-resource settings. The use of AI programs in healthcare has also raised several legal and ethical concerns, such as liability, privacy, and the potential for AI to replace human healthcare providers [217]. Another concern involves the overreliance on AI programs and models, which can lead to a reduction in critical thinking and clinical judgment among healthcare providers. This negative aspect affects both the professionals and the patient outcomes.
Ensuring robust data protection laws is of paramount importance in the era of big data, particularly in safeguarding the privacy of the patient. This situation raises concerns regarding the adequacy of existing regulations. It is imperative to address the shortcomings by encompassing health-related data that falls beyond the purview of current acts. Proactive measures are necessary to ensure comprehensive protection for health data, regardless of the entities involved in its collection, storage, and processing.
The sharing and regulation of disease-related data across multiple databases pose significant challenges due to the presence of personal information in patient records. This presents a complex landscape for software developers, who must navigate confidentiality regulations that can impede the development of AI. Privacy, in particular, is an important concern when dealing with health service data, as it represents the most private and personal information about individuals. Respecting confidentiality becomes an essential ethical principle in healthcare, intertwined with the autonomy, personal identity, and overall well-being of the patient [218].
On a different note, AI systems lack the empathy and personal touch that human healthcare providers can offer, which represent important aspects in patient care and satisfaction [13]. The role of human care providers extends beyond technical expertise. They engage in effective communication, and build trust with their patients. Instead of viewing the potential of intelligent artificial systems as replacements for human healthcare specialists, it is more appropriate to recognize the value of humans collaborating with these systems. The potential lies in integrating AI systems into healthcare workflows as tools to augment and enhance the capabilities of healthcare professionals.
Another three challenges of AI in healthcare include the black box problem, overfitting, and regulatory approval. A black box problem occurs when deep learning algorithms are unable to explain how their conclusions are reached. In the past, it was impossible to determine what imaging features were used in a process, how these were analyzed, and why the algorithm reached a particular conclusion [219]. Although the model could be simplified into a straightforward mathematical relationship linking symptoms to diagnoses, it is important to acknowledge that the underlying process may involve complex transformations that clinicians, and particularly patients, may struggle to comprehend. However, it is worth considering that the need for a complete understanding of the "black box" may not be necessary. In certain cases, positive results from randomized trials or other forms of testing could be sufficient evidence to demonstrate the safety and effectiveness of AI. While the internal workings of AI algorithms may remain complex and difficult to interpret, the focus can shift towards evaluating the overall performance and outcomes achieved through empirical validation.
Overfitting occurs when AI algorithms trained on one dataset have limited applicability to other datasets [220]. In this case, the algorithm has learned the statistical variations in the training data, rather than broad concepts required to solve a problem. The key determinant of overfitting is the overtraining of an algorithm on a specific dataset and several factors influence the likelihood of overfitting, including the size of the dataset, the extent of heterogeneity within the dataset, and the distribution of the data within the dataset.
Regulatory approval will pose a challenge for new AI algorithms. Medical AI, such as drugs and medical devices, require the FDA and regulation of other decisional organizations [221]. Due to the black box problem and overfitting, evaluators have difficulty understanding how algorithms work and whether their performance can be generalized to other datasets. AI tools are classified by the FDA based on three criteria: the risk to patient safety, predicate algorithms, and human input. In cases where algorithmic risks are high, such as diagnostic tools where a misdiagnosis would have severe consequences and where human input is minimal, premarket approval is conducted, which requires solid evidence that the tool is safe and effective from both non-clinical and clinical studies [6].
In addition, AI in medicine is still in the early stages of development, and there are limitations in terms of capabilities and accuracy of AI algorithms. While AI can analyze vast amounts of data quickly and efficiently, it may not be able to match the clinical expertise and intuition of human healthcare professionals. There may also be limitations in terms of the ability of AI to adapt to new situations and unexpected events, which can be critical in medical emergencies.

Implications for Practice
The progress of science and technology has sparked a notable increase in the utilization of AI and other ML techniques in modern medical practice [222]. AI integration into healthcare has become an essential catalyst for advancements in medical diagnoses and healthcare innovation in the era of 4.0. With the aid of medical AI technologies, medical specialists now have access to algorithms and programs that enable them to analyze patients' signs and symptoms, facilitating a deeper understanding of symbolic illness models and their interconnections.
Researchers in the field of AI have dedicated significant attention to diseases that are the leading causes of global mortality. It is projected that by 2030, chronic diseases will account for 80 percent of human lives lost worldwide, imposing substantial disease burdens on a global scale [223]. Consequently, researchers are leveraging cutting-edge technologies in the pursuit of early diagnoses and effective treatment approaches [224].
AI can assist a medical specialist by reducing the time spent on a diagnosis, allowing them to allocate more time to the patient's treatment. Additionally, AI enables medical personnel to proactively identify potential medical errors by extracting precise data [205] (Lee & Yoon, 2021).
The active involvement of patients in the medical care process plays a vital role in ensuring a disease diagnosis and promoting an effective treatment. For instance, in the case of anticoagulant therapy for stroke patients, an AI platform increased treatment adherence by 50% [225].

Future Directions
There is a clear transformation taking place in the field of medicine as AI continues to make its mark, modernizing various traditional medical components. The constant advancement of AI in this domain ensures the ongoing development of algorithms that can provide accurate and reliable diagnoses without liability concerns. To enhance the quality standard of AI algorithms, input data need to be combined with pattern recognition that offers valuable insights into the future. Predictive diagnostics will play a role in authorizing insurance claims, shifting the focus towards illness prevention rather than solely treatment. Patients can expect a same-day diagnosis, authorization, and treatment facilitated by interconnected AI systems across clinics and insurers.
Furthermore, AI will contribute to the integration of treatment options across different healthcare areas. As data-driven therapy continues to rise, the boundaries between medical disciplines are gradually merging, leading to the integration of comprehensive healthcare services.
As AI continues to advance, there is potential to enhance the efficiency of processes throughout an extended public health continuum. This advancement could enable the implementation of personalized prediction and prevention approaches that can be tailored to individual needs and applied across different populations. Such an approach has the power to significantly expand the scope of public health, with the involvement of various organizations beyond traditional public health institutions.
The widespread implementation of AI in healthcare necessitates increased data sharing. However, certain stakeholders exhibit reluctance to share their data with other parties due to concerns regarding the security of sensitive personal or commercial information. Consequently, healthcare competition and antitrust laws must adapt to comprehend the nuances of big data and AI.
The ideal level of trust required between clinicians and AI systems for making accurate and reliable clinical decisions remains uncertain. Additionally, the connection between optimal trust in AI systems and their design attributes is yet to be determined. Addressing person-specific factors, such as significant variability associated with aleatory processes and the evolving capabilities of AI, is crucial in analyzing the problem.

Contribution to Literature and Limitations
The purpose of this research is to disseminate information and enhance overall awareness regarding the utilization of AI in the healthcare sector. The aim is to facilitate the implementation of prospective decision systems and enable an early prognosis for patients. Specifically, we sought to determine if there are broader issues associated with emerging technologies beyond healthcare service transformation. The following are key contributions of this review: An overview and background of AI technology are provided to enhance the comprehension of cutting-edge concepts.
The context of AI in medical systems is explored, accompanied by a detailed discussion on ethical, legal, and trust-related concerns. This analysis aims to bolster public confidence in AI.
An examination of the reliability and utility of AI technology in healthcare applications is conducted.
After assessing the challenges and opportunities arising from the extensive integration of AI in healthcare, potential areas for future research are identified. These areas highlight avenues for further exploration and investigation.
Nevertheless, it is important to acknowledge the limitations of our research. Primarily, our study concentrated solely on a select few medical specialties within the vast realm of AI applications. This decision was driven by the authors' intention to remain within the boundaries of their own field of expertise.
Additionally, it is worth noting that our research takes the form of a narrative review, primarily due to the heterogenous nature of the included studies. In order to provide quantifiable data, further investigations such as systematic reviews and meta-analyses are required to yield quantifiable data and enhance the level of evidence in this particular subject matter.

Conclusions
In this review, we have conducted an analysis of the applications and impacts of artificial intelligence and machine learning in healthcare infrastructure. We have explored the diverse uses of AI in the medical sector, including areas such as diagnoses, prognosis research, and development. The review highlights the significant contributions that AI systems have made in healthcare by enabling machines to emulate human-like behavior and exhibit intelligent capabilities. This paper explores the benefits but also the challenges associated with integrating AI on a large scale in healthcare, and examines the ethical, legal, trust-building, and future implications of AI in the healthcare domain. Limitations of AI systems include the need for high-quality data, the potential for algorithmic bias, ethical concerns, and limitations in the capabilities and accuracy of AI algorithms. The insights presented in this paper aim to benefit the research community in developing AI systems tailored to healthcare, taking into account all the crucial aspects.
However, it is important to recognize the fact that our research focused solely on a limited number of medical specialties, a choice made to stay within the area of expertise of the authors. Furthermore, our research adopts a narrative format; additional investigations such as systematic reviews and meta-analyses are needed.
It is important to understand that we are still in the early stages of regulating the responsible design, development, and utilization of AI for healthcare, as the field is evolving rapidly. Nevertheless, it is our responsibility to conscientiously consider the ethical implications of implementing AI and to provide appropriate responses, even as the ethical landscape continues to evolve. As we continue to uncover their capabilities, AI systems have the potential to reshape healthcare delivery and improve patient outcomes.

Conflicts of Interest:
The authors declare no conflict of interest.