Artificial Intelligence‐Based Medical Sensors for Healthcare System

The aging population and the prevalence of infectious diseases have impacted the traditional medical order, significantly increasing the burden on healthcare and adversely affecting the socio‐economic system. Medical sensors based on artificial intelligence (AI) provide new ideas for modern medical data collection to monitor the health status of individuals and environmental changes. Meanwhile, with the aid of AI algorithms, the big data processing capabilities of sensor systems have been greatly improved, further realizing early predictions and timely diagnoses. In this paper, a brief overview is offered on the development status of AI‐enabled medical sensors for off‐body detection, near‐body monitoring, disease prediction, and clinical decision support system, and the ongoing challenges and future prospects to move from concept to implementation are discussed. In the foreseeable future, breakthroughs in the combination of medical sensors and AI algorithms are expected to pave the way for early detection and clinical decision support and improve the accuracy and efficiency for disease diagnosis.

[33] For example, Juang et al. used LSTM and RNN algorithms to analyze and learn the patient's historical medical data and clinician's decision, so that the proposed clinical decision support system (CDSS) can evaluate the patient's daily health status and provide comprehensive medical care recommendations. [34]AI algorithms can improve the sensitivity from 26.44% to 80.84%, meanwhile, yield 99.65% accuracy and 99.95% specificity.Through integrating this remarkable method into the existing hospital information systems, the developed CDSS is able to continuously monitor health status of patients, improve medical care quality and patient satisfaction.Therefore, the combination of accurate sensors and enhanced AI algorithms allows comprehensive information on disease characteristics through a sufficient number and variety of training samples.Systematic clinical assessment of health conditions can then be performed from multiple perspectives to improve the accuracy and efficiency of diagnosis.
Medical data is typically rich, rapidly growing, and relatively complex in structure.Machine learning (ML) techniques can combine medical datasets from millions of patients, such as diagnostic profiles, imaging records, and wearable information, to analyze the internal structure of the ocean of medical big data, identify patterns of disease conditions, and overcome the gen-eral limitations on access to local datasets.Furthermore, the nextgeneration healthcare system supported by big data shifts from a centralized hospital-based mode to a parallel mode of monitoring at home, screening and detection at point-of-care testing (POCT), and monitoring during hospitalization, meanwhile, achieves doctor-patient interaction and data transferring via the cloud to ease healthcare resource crowding and facilitate personalized healthcare (Figure 1).Ultimately, systematic health status assessment from comprehensive individual information is used for clinical applications to achieve improved data processing capabilities and resource optimization in healthcare.In this perspective, we present the latest advances in AI in clinical diagnosis from three perspectives on off-body detection, near-body monitoring, disease prediction and CDSS (Figure 2).The challenges and opportunities of AI in personalized medicine in the future are deeply considered and discussed.

Artificial Intelligence in Off-Body Detection
Off-body detection can be divided into biochemical detection and physical detection.Biochemical detection uses portable medical liquid and gas sensors to analyze biomedical markers in biofluids (blood, saliva, urine) and breath to determine an individual's state of health.How to avoid the interference of a large number of unrelated compounds and achieve accurate and effective disease diagnosis is a great challenge for biochemical detection.As for physical detection, the main part is to detect changes in the structure of diseased tissue and blood perfusion through images.However, current image-based diagnosis mainly relies on the experience of doctors and is less efficient and accurate.Compared with traditional disease detection methods, the novel emerging strategies involve the use of AI to collect datasets from medical sensors and images can effectively address the abovementioned shortcomings in traditional methods.

Off-Body Detection via Liquid Sensor
[37][38][39][40][41][42] However, interference from the irrelevant compounds, small sample volumes, and dilution of biomarkers are some of the challenges faced by modern medical sensors.Medical liquid sensors combined with AI algorithm can effectively avoid these problems.Shin et al. detected exosomes in hu-man plasma for early-stage lung cancer diagnosis by surfaceenhanced Raman spectroscopy (SERS) based on deep learning (Figure 3A). [43]They separated exosomes from human plasma samples and collected the SERS signals through a gold nanoparticle (GNP)-coated plate.Thereafter, they explored the features of plasma exosomes using deep learning and figured out the similarity of them, without learning insufficient human data.The supervised model trained with SERS signals successfully classified the exosome data into two clusters and predicted lung cancer patients and healthy controls with an accuracy of 95% and 90.7%, showing the potential of deep learning-based SERS techniques for lung cancer diagnosis as a routine prescreening tool.However, the analysis of blood extract biomarkers is invasive and requires complex pre-processing steps.Owing to the advantages of miniaturization, high throughput, and automation, microfluidic devices are popular in clinical diagnosis, which can provide comprehensive health information with small sample requirements.Braz et al. applied a single-response microfluidic e-tongue to obtain impedance data of saliva samples for oral cavity cancer diagnosis.Data were processed with multidimensional projection techniques, non-supervised, and supervised ML algorithms. [44]or the 27 individuals, distinguishing between healthy people Figure 3. Medical sensors for off-body detection.A) Schematic illustration of circulating exosome analysis to diagnose lung cancer based on deep learning.Reproduced with permission. [43]Copyright 2020, American Chemical Society.B) Schematic representation of the SiNW FET sensors to classify disease breathprints by ANN.Reproduced with permission. [58]Copyright 2016, American Chemical Society.C) Schematic of the DeepCOVID-XR to Detect COVID-19 on chest radiographs trained and tested on a large U.S. clinical dataset.Reproduced with permission. [67]Copyright 2021, Radiological Society of North America.
and those with floor of mouth or oral cavity cancer could only be achieved by supervised learning.Using the SVM with radial basis function kernel and random forest, for the binary classification (YES/NO for cancer), the accuracy is above 80%.While, for ternary classification (floor/cavity cancer, and control), the accuracy is reduced to 70%.With the help of a large number of samples, the accuracy values are expected to increase.The combination of microfluidics and AI shows huge potential in personalized medicine.47] Kim et al. introduced a urinary multimarker biosensor combined with two common ML algorithms for precise screening for prostate cancer (PCa).The biosensor contains four biomarker channels, which can bound to antibodies of each biomarker (PSMA, ENG, ERG, and ANXA3). [48]Through measuring a drop of natural voided urine from healthy individuals and PCa patients, the biosensor with single biomarker analysis has an average accuracy of 62.9%.Then, the two ML algorithms, random forest and neural network, were adopted to analyze the relationship between clinical state and sensing signals.With the increase of the number of biomarkers, the screening performance of the two algorithms improved monotonically.Through measuring the same batch of samples under the best combination of biomarkers, the ML algorithms can screen for PCa with more than 99% accuracy.Due to the advantages of simplicity and accuracy, multimarker biosensor approach has the potential to change the current PCa screening paradigm and impact the human healthcare system.

Off-Body Detection via Gas Sensor
[51][52][53][54][55] However, in order to bring breath analysis to the next stage to meet the unmet needs in modern clinical practice, there is an urgent need to improve the method of breath analysis so that it cannot only promote disease diagnosis, but also classify the conditions of diseases.Wang et al. applied ANN model to optimize the detection parameters of Si nanowire field effect transistors (SiNW FETs) device to achieve high selectivity. [56]The selectivity can achieve specific volatile organic compounds (VOCs) detection in single or multiply component environments as well as the estimation the compound concentrations.Through using one SiNW FET sensor instead of sensors array which is the common case in traditional electronic nose, up to 11 VOCs as well as their binary and ternary combinations can be precisely recognized.Especially in complex environments such as multicomponent mixtures with similar physical/chemical properties, the developed method shows excellent selectivity toward the targeted VOCs.Additionally, with the help of ANN model, the concentration of VOCs can be estimated with high accuracy.Therefore, the method of combining AI with sensors can be extended to more real-world sensing applications.Based on this method, Shehada et al. successfully developed an SiNW FETs array in conjugation with ML method to detect and discriminate between disease breathprints, such as gastric cancer, lung cancer, asthma, and chronic obstructive pulmonary disease (Figure 3B). [57,58]The fabricated sensors are optimized by a training dataset which is the relationship between the selectivity and the various disease breathprints.Using breath samples from patients and healthy volunteers, the accuracy of the sensor in discriminating between cancer and control conditions is above 80%.What is more, through coating with molecularly modification, the sensor could even distinguish the different stages of the cancer.This method can provide valuable diagnostic and treatment information, while reducing patient's discomfort, treatment delays, and financial costs.Exhaled sensing technologies have potential capabilities in the field of real-time tracking and monitoring of epidemics, and emergency medical alarms, both of which are necessary in the current COVID-19 outbreak.Shan et al. developed a noninvasive method to detect and follow-up people who were at-risk or had an existing COVID-19 infection. [59]A breath device composed of a nanomaterial-based hybrid sensor array with multiplexed detection capabilities was proposed to detect specific biomarkers from exhaled breath, achieving rapid and accurate diagnosis.An exploratory clinical study was carried in Wuhan, China, during March 2020, including 49 COVID-19 patients, 58 healthy individuals, and 33 controls with non-COVID lung infection.By training the device signals of the three groups, the discriminant factor analysis (DFA) model exhibits 90% and 95% accuracy respectively in distinguishing COVID-19 patients and the control group without COVID lung infection.The combination of the proposed device and AI algorithms will lower the burden on hospitals and prevent other patients from fearing medical care.Furthermore, the proposed method can be regarded as a personalized platform that can be adapted to other disease infections with appropriate modifications to the artificial intelligence.Thus, this approach could be utilized as a diagnostic tool in the event of a new disease outbreak.

Off-Body Detection via Image
[62][63] AI has been applied on different scale databases with probabilistic and statistical methods for imaging field.Imaging feature processing and ML-based classification or prediction have great potential in helping radiologists make diagnoses as accurate as possible to reduce diagnostic time and cost.Additionally, it can help radiologists consider regions of interest to predict cancer that would otherwise be overlooked.Esteva et al. used deep CNNs to automatically classify skin lesions under binary classification. [64]The computational method has shown a comparable ability to dermatologists in tests for the identification of the most common cancers and the deadliest skin cancer.Additionally, CNN is trained directly by the image pixels and disease labels, which can avoid extensive preprocessing compared with traditional classification methods.This fast, scalable method allows medical and patients to detect skin lesions earlier and, therefore, has a potential impact on substantial clinical.To extend the method, Fraiwan et al. classified the skin lesion images into seven categories. [65]The high accuracy can reduce the burden on specialists and the easy implementation can promote the practical application.However, the imbalance of the existing dataset, the small number of images, and large number of classes reduce the overall accuracy of the method, which is also the focus of future research.Existing challenges of lung cancer include inter-grader variability and high false-positive and false-negative rates.Ardila et al. proposed a deep learning algorithm that used current and prior computed tomography volumes of a patient to predict lung cancer risk. [66]They constructed a 3D CNN model and then trained a CNN region-of-interest (ROI) model to detect 3D cancer candidate regions.Based on this, they eventually developed a cancer risk prediction model to assign a case-level malignancy score.For 6716 National Lung Cancer Screening Trial cases, the model achieves a state-of-the-art performance of 94.4% of the area under the curve (AUC), which is similar to an independent validation set of 1139 cases.Furthermore, without prior computed tomography imaging, the model has a better performance than all six radiologists with absolute reductions of 11% in false positives and 5% in false negatives, showing huge potential of deep learning models in accuracy, consistency, and adoption improvement for lung cancer screening.In addition to cancer detection, images are also a promising detection method for the currently popular COVID-19.However, previous algorithms for identifying COVID-19 have problems, such as small datasets and poor validity.DeepCOVID-XR, as a deep learning AI algorithm, was used to detect COVID-19 on chest radiographs based on reverse-transcription polymerase chain reaction test results (Figure 3C). [67]Each image in the dataset is first preprocessed to generate four separate images and fed into six previously validated CNN architectures.Through comparing the weighted average of these individual CNN predictions with an output threshold, images can be binary classified into COVID-19 positive or negative.In a test of 2214 images, the proposed algorithm has an accuracy rate of 83% and AUC of 0.9, meanwhile, in a test of 300 random images, it has an accuracy rate of 82% and AUC of 0.88, which shows the similar performance with experienced thoracic radiologists.In addition, other clinical data (e.g., vital signs, laboratory data) are planned to incorporate into the algorithm to further enhance its performance and adapt it for predicting the risk of clinically significant outcomes in patients with confirmed COVID-19.

Artificial Intelligence in Near-Body Monitoring
Off-body detection presents advantages in terms of experimental control and reproducibility.However, it also poses certain drawbacks, including cumbersome operation and intermittent data collection.In contrast, near-body monitoring through wearable sensors enables non-invasive, real-time, and long-term monitoring of health information.[70][71][72][73][74][75][76][77][78][79][80][81] Currently, specific applications based on nearbody monitoring include disease, motion, and mental status monitoring.Further advances in near-body monitoring are likely to focus on the development of multiplexed biosensors to monitor several biomarkers levels in real time and multiparametric cross-referencing among multi-sensor parameters.

Near-Body Disease Monitoring
[84][85][86] Diabetes is one of the most common chronic diseases, which need to check the blood glucose level of patients daily for continuous management.Lee et al. integrated monitoring and treatment into one wearable device based on functionalized chemical vapor deposition (CVD) graphene. [87]The solid-state Ag/AgCl counter electrodes were used to enhance the biochemical sensors' electrochemical activity, sensitivity, and selectivity for the human sweat biomarkers detection.When high glucose concentration was detected, a heating component could be triggered to enable feedback dosing.They then fabricated the monitoring and therapy device based on a thin polyimide (PI) substrate, which makes it easier to integrate and industrialize (Figure 4A). [88]The two loadable drugs could realize slow suppression and fast regulation of the blood glucose levels respectively.The novel integrated system provides important advances in painless and stressless diabetes treatment.In addition to diabetes detection, the level of Levodopa (L-dopa) in sweat can be used for Parkinson's disease drug dosage monitoring.Xiao et al. reported an electrochemical sensor to monitor the concentration of L-dopa in sweat based on metal-organic framework with the integration of enzymes. [89]he sensor was yielded by loading tyrosinase onto the composite, which was synthesized through the in situ growth of ZIF-8 nanoparticles on the surface of graphene oxide (GO), forming the Zeolitic imidazolate framework/graphene oxide (ZIF-8/GO) composites.Communication between the medical sensor and a smartphone application for real-time monitoring is established through the integration of a wireless electronic circuit.The platform has high sensitivity and good stability with a wide linear response range from 1 to 95 μm and low LOD of 0.45 μm, showing great potential for continuous and noninvasive drug monitoring and health management.Heart disease is the leading cause of death for people worldwide.However, the evolution of electrocardiogram (ECG) signals from the human body are too weak and low-frequency to be detected, which places a higher requirement for data acquisition and processing.The combination of wearable device and AI will promote our understanding of cardiac physiology.Li et al. reported a wearable ECG acquisition device to detect ECG signal from human body, and an improved ResNet-based ECG algorithm combined with AI model to classify ECG signal accurately and rapidly. [90]The system can diagnose seven common cardiac arrhythmias in real time with an accuracy of 98.3%.Moreover, it offers access to cardiac science-related knowledge, enabling users to obtain a more comprehensive understanding of ECG-related information.This serves to enhance their comprehension of the prevention of cardiovascular and cerebrovascular diseases, thereby significantly improving patients' self-preventive abilities.

Near-Body Motion Monitoring
[93][94][95][96][97] In this regard, reliable tools for convenient signal acquisition and efficient data processing are important and highly desired.Novel skin-interface devices, such as electronic skin, fiber sensor, and hydrogel sensors, have attracted great attention owning to their potential applications in motion monitoring.Lin et al. reported an all-fiber motion sensor (AFMS) with excellent flexibility, breathability, biocompatibility, and motion recognition ability that, combined with AI algorithms, can be applied to identify throat motion states with above 85% accuracy (Figure 4B). [98]In practical application, under the flexibility guarantee of AFMS, the degree of sensor bending determines the signal amplitude and the motion frequency determines the signal frequency.In order to solve the indistinct signal features interference in coughing detection, three ML methods (extreme gradient boosting (XGBoost), k-nearest neighbor (KNN), SVM) were introduced to effectively classify throat motions and noise signals.The design of this wearable sensor combined with AI has great significance for more efficiently and accurately detecting human motion and developing nextgeneration healthcare systems.The combination of physical and biochemical detection enables comprehensive profiling of human subjects engaged in extended indoor and outdoor physical activity and permits real-time evaluation of the subject's physiological state.Gao et al. proposed a fully integrated wearable sensor array capable of performing multiplexed in-situ analyses, which can selectively detect sweat metabolites and electrolytes through biochemical methods, while also responding to skin  [88] Copyright 2017, American Association for the Advancement of Science.B) Schematic representation of the radial anisotropic porous silver fiber-based ultralight sensing material to measure body motion as a sensor.Reproduced with permission. [98]Copyright 2022, Springer Nature.C) Schematic of a new design for explainable AI used in stress prediction based on physiological measurements.Reproduced with permission. [108]opyright 2022, BioMed Central.temperature changes by physical means. [99]Integration of a plastic-based sensor in contact with the skin and a silicon integrated circuit on a flexible circuit board enables complex signal processing and overcomes the technological gap between signal transduction, modulation, processing, and wireless transmission in wearable biosensors, resulting in a wearable biosensor that is fully functional and capable of personalized diagnostic and physiological monitoring.For motion monitoring, sensors may be accidentally scratched and/or cut during the required prolonged use or wear, thereby destroying their sensing capabilities.[102][103] Khatib et al. introduced an electronic skin with self-repair capability to detect temperature, pressure, and pH levels sensitively under complex environments. [104]The soft electronic device integrated a network of neuron-like nanostructures for damage detection and self-monitoring, as well as an electric heaters for selective self-healing.This remarkable electronic platform lays the foundation for the novel self-healing electronics and improve the sustainability and durability of soft electronic devices.Recently, Xia et al. fabricated a hydrogel strain sensor with high flexibility, self-adhesivity, self-heal ability, and conductivity for human motion monitoring. [105]hrough attaching the sensor to various body parts, the output resistance signals can effectively indicate flexion and relaxation of wrist, elbow, neck, and knee joints.Especially, the sensor can monitor very subtle activities, such as speaking, breathing, and pulse.The hydrophobic association cross-linked network forms a hybrid network hydrogel through network interpenetration for excellent self-healing capability and reproducible adhesion properties, which is suitable for motion monitoring and has significant impact on extending the life and durability of wearable devices.

Near-Body Mental Status Monitoring
The monitoring and evaluation of mental status are difficult, because it is hard to recognize or quantify.Using wearable devices to moment-by-moment quantify a person's daily life is a possible solution to the problem for monitoring and recognition.
Combined with AI, the mental statuses of patients can be intervened in time to avoid tragedies. [106]Riera et al. described a stress monitoring system, where the analyzed data came from the fusion of electroencephalography (EEG) and electromyography (EMG). [107]EEG was used to extract the emotional state information which is most related to stress.EMG was used to improve the robustness of the monitoring system.The data fusion method can effectively improve the classification rate from 79% to 91.7%, which is an excellent performance for a real-time system.While significant progress has been made in the development of mental status monitoring models, little work has been done in the development explainable of AI for mental health.ML, as a branch of AI, can play great advantages in multi-biosensor data fusion aspect and data interpretation.Jaber et al. developed an AI system design to explain mental health on the basis of the physiological data recorded by wearable devices (Figure 4C). [108]irst, the shapley additive explanations (SHAP) model is used to calculate contribution of each physiological factor to the overall stress probability.Then, the balanced random forest classifier was chosen to predict stress because of the reliable results and strong ability to handle unbalanced datasets.Finally, the system will generate a report showing the measured values and normal ranges for each component of the stress assessment.The F 1 score of the system is 0.78 and the accuracy is shown to be comparable to the state-of-the-art.Furthermore, ML can be used for anxiety monitoring.Zeng et al. reported epidermal electronics systems (EES) to simultaneously detect multiple physiological signals, which adopted ML algorithms to classify and predict mental fatigue levels. [109]The EES consists of two modules: the first module is attached to chest for ECG and respiration rate monitoring; the second module is attached to one palm for galvanic skin response detecting.Then, three kinds of ML algorithms (SVM, KNN, DT) extracted features from the induced signals and built predictive model to detect fatigue status.Based on six types of physiological characteristics, the prediction accuracy can achieve up to 89%.This system, which is relatively simple to be fabricated, provides a powerful strategy for the further development of epidermal multifunctional sensors.Compared with mental fatigue, generalized anxiety disorder is a more severe condition, which is much harder to monitor and evaluate.Jacobson et al. used passive wearable sensor data gathered from past week combined with a ML model to predict GAD symptom severity. [110]The data is available from the National Health and Nutrition Examination Study.They proved that there existed strong discriminant validity between risk score calculated from the ML model using one week data and GAD symptom severity.In terms of clinical utility, the risk score could accurately detect the patients with GAD symptom.Due to the poor diagnosis rate and healthcare, this study based on AI algorithms provides an opportunity for noninvasive evaluation of GAD symptoms.

Artificial Intelligence in Disease Prediction and Clinical Decision Support Systems
Healthcare providers are searching for lower-cost and more rapid ways to deliver healthcare services.The traditional hospitalcentered healthcare system focuses on diagnosis first and treatment later.[113][114][115][116][117] On one hand, AI algorithms have the ability to merge deep analysis with powerful predictive capabilities, providing fast disease predictions for future data through large amount of medical data processing and model training.On the other hand, CDSS helps the decision-makers and healthcare systems to improve the information, insights, and environments approaching way.For example, a CDSS combined with AI can take into account individual variability in environment, lifestyle, and genetic make-up to achieve the mentioned targets.In a real-life environment, the application of the tool will promote the ability of health assessments to predict and detect specific diseases, distinguish disease subcategories and genetic characteristics, and monitor disease progression and treatment processes.Representative works for disease prediction and CDSS based on AI are presented in Table 1.

Machine Learning Algorithms in Disease Prediction
][120] In this situation, the ML algorithms are typically supervised.Supervised learning uses sample inputs of known classes to make predictions.The data for model building is given in advance and is fixed, which has the same statistical properties as the final data used.The following ML algorithms are commonly used in clinical management and disease prediction, including regression-based ML algorithms, SVM, and CNN.
Regression-based ML models are able to use one or more variables to predict a continuous output variable by linear or nonlinear function fitting.It is primarily used for time series modeling and for determining causal relationships between variables.

Zhai et al. developed a logistic regression algorithm based on
Electronic Health Record to combine physiologic and/or laboratory measures for early clinical deterioration prediction. [121][124][125] Considering the correlations based on wellknown statistical principles, regression-based models offer a simple and intuitive approach to medical data classification and prediction, but require more strict assumptions.
SVM, as a supervised ML algorithm, can divide multiclass problems into several binary problems to solve the "multiclassification" challenge and one-to-one predictive challenge. [126]To  [ 118] Flexible inductor-capacitor (LC) resonant sensor ATmega328 hardware Deep ANN 0.946 accuracy Prediction of tissue regeneration [ 119] Inertial-sensor Wearable gait device Random forest 0.9375 accuracy Identification of patients with osteopenia and sarcopenia [ 120] 16 clinical indicators (temperature, systolic blood pressure, etc.) Weka 3.6.8software Logistic regression 0.849 sensitivity 0.859 specificity Prediction of intensive care [ 121] Surface EMG and inertial sensors MATLAB software Support vector regression 0.059 mean square error Quantitative assessment of muscle spasticity [ 125] 15 Trigno Wireless sensor MATLAB 2017a software SVM 0.9669 accuracy 0.9824 sensitivity 0.9603 specificity Prediction of human tic [ 126] 4 clinical indicators (GSH, protein content, etc.) R platform (v3.6) software SVM 0.9757 AUROC Prediction of the severe/critical symptom of COVID-19 [ 127] Photoplethysmography sensor TensorFlow(v.2.2.0)software CNN 0.9932 accuracy 0.9932 F1-score Recognition of human activity [ 128] Wearable ECG patch sensor iOS APP/IoT hardware CNN 0.9496 accuracy Prediction of cardiac disease [ 129] 3 tumor marker indicators (CYFRA21-1, CEA, and CA-125) One-vs-rest software CNN 0.84 accuracy Prediction of NSCLC [130] Wearable commercial sensors Integrated development environment platform SVM and LSTM Correlate different health algorithms and models for the first time DSS oriented on health prediction [ 132] Wearable medical sensors TensorFlow software DNN 0.9925 accuracy 0.9803 specificity CDSS for CKD [ 133] Glucose monitoring Glooko platform AI Non-inferior to intensive insulin titration provided by physicians DSS for insulin dose optimization [ 134] divide the input dataset into potential classes, SVM constructs one or a set of hyperplanes in high-dimensional spaces.The margin distance, which is the distance between the hyperplane divider line and support vectors, indicates the accuracy of classification results.For problem such as classifying symptoms of patients with clinical indicators, SVM provides an attractive solving method.Sun et al. proposed SVM based on the combination of features for severe/critical COVID-19 cases prediction. [127]In this study, 336 patients infected COVID-19 in Shanghai as of March 12, 2020 were divided into training and test datasets.220 clinical and laboratory observations or records were also collected to identify clinical indicators associated with severe or critical symptoms and develop prediction model.A total of 36 clinical indicators were identified, including thyroxine, immune-related cells, age and so on.The optimized combinations of characteristics reach 99.96% and 97.57% of the area under receiving operating curve (AUROC) in the training and test datasets, respectively.The study demonstrates that SVM is robust and effective in disease prediction, showing great potential in screening the severe/critical cases early and saving medical resources.However, it is difficult to train SVM models with large training samples because it has high computational complexity and will consume a large amount of computer memory and computation time.
CNN is a kind of feed-forward neural network with convolutional calculation and deep structure.It is one of the representative algorithms of ML and has the ability of representation learning. [128]CNN is mainly composed of convolutional layer, pooling layer, and fully connected layer.The convolution layer contains multiple convolution kernels and has the ability to learn the local characteristics of the input dataset.Pooling layers are used to reduce the dimensionality of the dataset.The fully connected layer is equivalent to the hidden layer in the traditional feedforward neural network, and only transmits signals to other fully connected layers.Lin et al. proposed an AI internet of things (IoT) system for ECG analysis and cardiac disease detection. [129] wearable ECG patch, as the front-end hardware, is used to detect the ECG signals.Based on the CNN algorithm, the big-data database storing each user's ECG signal can instantly mark abnormal ECG signals to achieve early disease prediction and realtime detection with an accuracy of 94.96%.The wearable patch includes a power management integrated circuit (IC), a level shifter, a 10-bits analog to digital converter, a bio-signal processor and a Bluetooth module.With the help of two silver chloride wet electrodes, the patch can be used for up to 24 h.The algorithm includes two segments: data pre-processing and CNN model, and can divide the arrhythmia into four categories: normal, atrial fibrillation, flutter, and ventricular fibrillation.The data pre-processing consists noise removal, baseline removal, and image generation.1D CNN is designed as four convolutional layers and three fully connected layers for feature extraction and classification.Similarly, Zhan et al. used CNN to develop a sensor-aided prediction and decision-making intelligent medical system (Figure 5A). [130]The patient's medical record can be obtained by sensor in real time, then it is converted into corresponding text description data.As a local connection network with automatic learning for feature extraction, CNN can semantically represent text and extract features, and represent words in  [130] Copyright 2021, The authors, published by MDPI.B) Schematic of a cloud-based CDSS employing DNN classifier for chronic kidney disease diagnosis.Reproduced with permission. [133]Copyright 2019, Elsevier.
text as continuous dense vectors in multidimensional space, and words with similar semantics correspond to similar word vectors.Through testing real-world non-small cell lung cancer (NSCLC) patient case samples, the results show that the intelligent medical system can achieve an accuracy rate of 0.84 in the diagnosis of NSCLC staging.This remarkable system can help evaluate the effect of patients' treatment and timely adjust the treatment plan for the next stage according to the patient's recovery.

Artificial Intelligence in Clinical Decision Support Systems
AI has been used widely in CDSS, which extracts insights from multiple sources of medical history records and diagnostic results to improve the accuracy and efficiency of information analysis at a deep level and prevent or minimis the medical errors. [131]CDSS represents a comprehensive and personalized decision support system that assists physicians in making clinical decisions more scientifically and efficiently, with the aim of enhancing the quality and efficiency of medical services, extending beyond its utility solely as a tool for disease prediction.
The adoption of AI in CDSS will support a considerable range of decisions, especially those under uncertainty.It will also discover the information hidden within clinical data, highlight relationships among them, and assess the consequences of proposed solutions.Massaro et al. proposed a decision support system (DSS) combined with AI algorithms and wearable sensors to predict individual health status. [132]The DSS consists of wearable sensors module, multi-dimensional 3D cubic module, and AI engine module.The wearable sensors module can obtain the patients' physiological data and then send to a cloud platform.The multi-dimensional 3D cubic module maps the acquired data into a 3D space, where three axes represent health risks related to pollution, emotional state, and physical activities.AI engine module combines SVM and LSTM algorithms to classify and cluster patients through pathologies, predict the evolution of pathology, assess cardiac and respiratory failure, and endow data analysis with geographic location stamp.After validating the basic functions of physiology prediction, this work is suitable for multi-dimensional health status risk mapping.Similarly, Lakshmanaprabu et al. proposed a novel cloud-based CDSS framework to predict and identify the presence and severity of chronic kidney disease (CKD) (Figure 5B). [133]The healthcare data of patients comes from three methods: sensor-based wearable IoT tools, dataset of University of CaliforniaIrvine repository, IoT devices.In terms of algorithms, deep neural network (DNN) is used to identifies CKD and determines the severity, meanwhile, particle swarm optimization (PSO)-based feature selection technique is adopted to remove noisy interference and exclude unwanted features.The DNN classifier, as proposed, achieves a predictive accuracy of 98.25% for CKD detection.This performance is further improved to 99.25% through the utilization of the PSO-based feature selection method.Furthermore, the enhanced classification accuracy is confirmed with higher specificity (98.03%), accuracy (99.25%),F-score (99.39%), and kappa value (98.40%).In addition to being used to detect diseases, CDSS can also provide support for disease treatment.Nimri et al. described an AI-based automated DSS which was used to adjust insulin dosage for youths with type 1 diabetes. [134]AI-DSS consists of a secured database and a data-processing algorithm module.The system uses the raw data from the database to detect patient dosing decision event which is characterized by the insulin pump and continuous glucose monitoring system.Meanwhile, it is able to detect the blood glucose mode: hypoglycemia, euglycemia, and hyperglycemia.Thereafter, the combination of event-based analysis and the patient's blood glucose mode can generate personalized treatment recommendations.The findings indicated that the use of the AI-DSS for optimizing insulin pump settings was not inferior to intensified clinical care delivered by trained physicians in terms of efficacy and safety.Furthermore, physicians who utilized the AI-DSS expressed a high level of satisfaction.The AI-DSS was deemed safe and the proportion of readings falling into the hypoglycemic range below 54 mg dL −1 in the AI-DSS arm was statistically equivalent to that observed in the physician arm.
To reveal health states in large amounts of information, CDSS will enhance preparedness by combining verifiable datasets from millions of patients' information.Therefore, privacy concerns surrounding medical data, particularly personally identifiable patient information, have received significant attention.It is becoming more and more important to develop privacy-preserving AI and methods which make AI models learn from datasets without compromising privacy.Bonawitz et al. proposed the federated learning (FL) method, which enables AI algorithms to gain experience from large amounts of data located at different sites. [135]ultiple sites can collaborate on model development without sharing sensitive data directly with each other, which is suitable for CDSS.

Conclusion
Benefiting from the development of AI-enabled medical sensors, the healthcare field are revolutionizing from acute, reactive, and preventative to pro-active healthcare.Diagnosis and clinical decision-making supported by AI and sensing technologies can be used for various physical and biochemical markers detection of diseases in a continuous and non-invasive manner as an integrated suite of innovative tool, providing a selective way for inpatient and remote patient monitoring.The medical sensors combining with AI technologies can be divided into three categories according to different functions.i) Off-body detection can be divided into biochemical and physical types for disease diagnosis, detecting biomarkers by liquid or gas sensors and identifying changes in body structures by images; ii) Near-body monitoring collects vital signs from skin by wearable sensors, which can be applied in disease, motion, and mental status monitoring; iii) Disease prediction and CDSS are able to provide accurate and ongoing medical decision-making in early diagnosis, improving well-being and quality of patient's life.
In recent years, the combination of medical sensors and AI algorithms has aroused widespread interests and made significant progress.Nonetheless, few new developments are commercialized and fully used in clinical applications.Challenges need to be overcome to increase the market and medical value of biosensors.For off-body detection, the concentration of health markers in biofluids and exhaled breath are too low to distinguish signals from noise by clinical sensors.Meanwhile, current in off-body devices have low level of automation, specificity, stability, and require complex sample pre-processing.Therefore, novel materials and structure are needed to improve the detection performance.Advanced fabrication technologies and new AI algorithms can be used in automated system integration for labor freeing and time saving.For near-body monitoring, large-area fabrication and interconnected lines layout increase the difficulty of designing and manufacturing wearable sensors.The quality of the interfaces between wearable sensors and target organs affects the efficiency of signal acquisition.Furthermore, power supply system and wireless sensing communication of wearable devices should be improved to ensure long-time operation of sensors and transmission of bulk data.For disease prediction and CDSS, most platforms focus only on health characteristics, ignoring potential disease signals, and cannot really serve as an early warning.Therefore, the program platform should deeply understand the medical needs and pain points to achieve specialization and practicality in parallel.Moreover, to improve the social acceptance of the platforms, the devices or apps need to be affordable and compatible with most mobile operating systems and smartphones from different manufacturers.Humanized interaction will also increase the ease-of-use and pervasiveness of platforms.For AI algorithms, it already has the ability to process large amounts of data quickly and relatively efficiently, such as voice, image, and video.However, challenges in data analysis and interpretation remain largely unaddressed.Researchers can endeavor to enhance the interpretability of algorithms by developing more transparent decision trees and linear models, in addition to creating more interpretable black-box models, including interpretable algorithms based on generative adversarial networks.Meanwhile, the absence of explicit criteria for algorithm selection may impact the interoperability of data and the portability of algorithms.Establishing technical standards represents an important approach to address this issue, as it can facilitate seamless integration and interaction between medical sensors and artificial intelligence algorithms produced by different providers, while enhancing the repeatability of algorithms.Furthermore, the privacy and security protection of sensitive medical information is not taken seriously.By leveraging encryption techniques and implementing authentication access control mechanisms, computational tasks can be performed without revealing sensitive data, and access to sensitive data and algorithms can be restricted, thus protecting data security.AI algorithms need to combine knowledge-based and non-knowledge-based models based on clinical guidelines to present evidence-based results, understanding and explaining each content like a human expert.Therefore, more robust algorithms should be developed and implemented to make sure the medical privacy and information security.
Even though there are challenges in commercialization and clinical application, AI-enabled medical sensors will still open up many new opportunities in health monitoring, disease diagnosis, and prediction.It is believed that with the combined efforts of researchers around the world, the combination of medical sensors and AI algorithms will provide a new platform for more efficient and accurate clinical decision-making, and has the potential to improve nearly each aspect of healthcare management in the future.

Figure 1 .
Figure 1.Schematic of current healthcare system and next-generation healthcare system based on big data for early diagnosis and healthcare facilitation.

Figure 2 .
Figure 2. Conceptual diagram of the combination of medical sensors and AI algorithms for off-body detection, near-body monitoring, disease prediction, and CDSS.

Figure 4 .
Figure 4. Medical sensors for near-body monitoring.A) Schematic of wearable sweat-based device for glucose, pH, and temperature monitoring.Reproduced with permission.[88]Copyright 2017, American Association for the Advancement of Science.B) Schematic representation of the radial anisotropic porous silver fiber-based ultralight sensing material to measure body motion as a sensor.Reproduced with permission.[98]Copyright 2022, Springer Nature.C) Schematic of a new design for explainable AI used in stress prediction based on physiological measurements.Reproduced with permission.[108]Copyright 2022, BioMed Central.

Figure 5 .
Figure 5. AI algorithms-based disease prediction and CDSS.A) Diagram of a CNN-based decision-making intelligent medical system for non-small cell lung cancer prediction.Reproduced according to the terms of the CC BY license.[130]Copyright 2021, The authors, published by MDPI.B) Schematic of a cloud-based CDSS employing DNN classifier for chronic kidney disease diagnosis.Reproduced with permission.[133]Copyright 2019, Elsevier.

Table 1 .
AI in disease prediction and clinical decision support systems.