Introduction

In 1956, an American computer scientist John McCarthy first introduced the term and principles of ‘artificial intelligence (AI)’ [1, 2]. The term AI is used to describe ‘machines’ capable of demonstrating cognitive functions that humans associate with other human minds such as ‘learning’ and ‘problem solving’ [3]. The foundation of AI is ‘machine learning’, a form of intelligence based on the compilation of a complex algorithm and software that mimics the human mind to decipher critical problems that include visual perception, speech recognition, and decision-making [4]. Similarly, ‘deep learning’ is described as a class of artificial neural networks that learn in a supervised and unsupervised manner [5] (Fig. 1).

Fig. 1
figure 1

Artificial intelligence and subtypes

Full size image

In health care, AI uses this ability to analyze medical data in disease prevention, diagnosis, patient monitoring, and development of new protocols [6]. Clinicians today are drowning in voluminous data, but AI gives the hope of easing the burden [7]. After the digitization of health-care data beginning in the mid-1960s until the incorporation of the electronic health record (EHR) in the American Recovery and Reinvestment Act 2009 [8], the increasing availability and progress of analytical techniques are opening new doors in health care [9]. AI is helping clinicians assess only clinically relevant information buried under massive amounts of data.

Neurological disorders comprise structural, biochemical or electrical abnormalities of the brain, spinal cord and nerves. With increasing population and aging, the burden of chronic neurological disorders has increased substantially despite a decrease in mortality due to stroke and other communicable neurological diseases [10]. In 2014, nine of the most common neurological disorders—dementia, stroke, epilepsy, Parkinson's disease, multiple sclerosis, migraine, and tension-type headache—cost the US economy nearly $789 billion [11]. In 2015, neurological disorders caused 250.7 million disability-adjusted life years (DALYs), which comprised 10.2% of global DALYs and 9.4 million deaths, and16.8% of global deaths [12]. Today, neurology faces multiple challenges in the field of diagnostic and management modalities. This ranges from simple issues like identification of healthy sleep patterns to more complicated issues like early detection and reduction in the duration of rehabilitation of acute ischemic stroke diagnosis of rare subtypes of epilepsy and prevention of sudden unexpected death in epilepsy (SUDEP), or considering multiple attributes of epilepsy and inter-observer variability of reading EEG. The vast amount of data compiled in neurology every year requires deep learning to structuralize the data to help neurologists in making an early diagnosis and improving care [13]. Artificial Intelligence has gained a lot of attention in the realms of detecting, diagnosing, and even preventing irreversible outcomes due to neurological disorders [13].

The founding principle of precision medicine is AI, which will eventually be part of neurological management. It is an emerging approach for disease treatment and prevention that takes multiple variables such as gene, environment, and lifestyle into account. It has the potential to work at an unprecedented rate by using massive computer power without any human programming [14]. The future of AI in every field is reassuring, especially in neurology from the prediction of outcomes of seizure disorder, grading of brain tumors, upskilling neurosurgical procedures, rehabilitation of stroke patients to smartphone apps monitoring patient symptoms and progress, the future seems promising [15]. During the last two decades, multiple gadgets equipped with artificial intelligence have been invented and researched to improve the functionality, diagnostic efficiency, and prognosis in patients suffering from neurological disorders. Few examples of such include the Apple Watch monitoring tremors and asymptomatic arrhythmia; iPad devices monitoring patient drawing capability with suspected movement disorders; CT scans and MRI equipped with AI to help radiologists diagnose complicated images; AI health applications to improve patient medication adherence; EpiFinder application to identify seizure types and epilepsy syndromes for triage [16]. A new horizon is emerging in neurology in the form of artificial intelligence, helping patients to improve their prognosis.

The primary objective of this narrative review is to highlight the emerging AI technologies that are creating new ways of neurological disorder management and improvement of patients’ overall functional outcomes. We have described various forms of AI used and available for use, as well as prior research, outcomes, benefits, and limitations of AI, effective accessibility and future of AI, keeping the current burden of neurological disorders in mind.

Discussion

Types of AI and prior research

In spite of the voluminous research helping to diagnose a complex spectrum of diseases, the translation into clinical practice has been challenging. Machine learning (ML) may be able to bridge the gap between translation of relevant clinical data and accurate clinical diagnosis [7]. Smart devices with AI technology, including smartwatches, smartphones, and tablets are being used by researchers not only to identify and stratify complex movement disorders [17] or arrhythmias including atrial fibrillation [18], but also to predict aspiration pneumonia in patients with swallowing difficulties secondary to stroke and dementia [19] and to improve medication compliance in patients on anticoagulation therapy [20]. Epileptologists are using smart devices with wrist annotated sensors to detect seizure activity [12, 21] and AI-enabled iPads to come up with a differential diagnosis and therapeutic approaches for rare epilepsy syndromes [22]. These studies could further pave a way to a better understanding of the pathophysiology involved in SUDEP [21]. Table 1 lists various types of AI technology and their clinical applications in neurological disorders.

Table 1 Types of artificial intelligence technology and their applications
Full size table

Different algorithms play an integral part in ML. ML algorithms such as random forest that use neuroimaging data in small sample data with high-dimensional parameter spaces are more stable than other algorithms and have been successful not only in classification of dementia, including mild cognitive impairment (MCI) [23] and Alzheimer’s [24], but also in other disorders including psychogenic non-epileptic seizures (PNES) [23, 25], Parkinson's disease [26, 27], and schizophrenia. In the same note, different algorithms have been used in the classification of similar disorders such as dementia [23, 25, 28]. Algorithms of the self-organizing map have been used by a group of researchers to identify distinctive phenotypes among concussive patients [29]. Deep learning algorithms are being used to predict the time since stroke onset (TSS) to help clinicians come up with better tools to guide stroke treatment [19]. EpiFinder algorithms help in diagnosing seizure in adult patients with spells [22]. These tools can help diagnose elusive disorders and prevent delay in diagnosis and treatment. Prior studies, utilizing the role of AI in neurological care, are mentioned in Table 2.

Table 2 Use of artificial intelligence and machine learning for management of neurological disorders
Full size table

Stroke

Stroke is the leading cause of disability and the fifth leading cause of death in the USA (102). Each year 795,000 Americans experience a new or recurrent stroke [30], and an estimated 24 billion dollars is spent annually on direct medical expenses [30]. Only 5% or fewer receive intravenous thrombolytic therapy in spite of the urgent need to administer it to preserve tissue in acute ischemic stroke [31, 32]. This can be due to lack of physician experience in administering thrombolytics, risk of 6% hemorrhage with thrombolytics, patients living in rural areas with limited resources, and strokes are unwitnessed or wake-up strokes [31,32,33,34]. Thus, there is an urgent need to streamline care and improve technology to solve this complex issue and bring down the rising costs [35]. ML has shown that it not only predicts the risk of recurrent stroke within 1 year after a TIA or minor stroke [36], but also predicts time since stroke onset (TSS) and is a better alternative than current DWI–FLAIR mismatch [37,38,39,40,41,42] in patients with unknown TSS (wake-up strokes or unwitnessed strokes); thus, it guides physicians to make better therapeutic approaches [19]. Smart devices with apps, using techniques like photoplethysmography, and handheld electrocardiograph recorders with greater accuracy are being used to check the heart rate and heart rate variability and to screen for asymptomatic atrial fibrillation which helps to prevent embolic stroke [18].

Epilepsy

Given its varied clinical manifestations, the rates of misdiagnosis are 26% in epilepsy centers and 20–40% in community settings. This often leads to unwarranted investigations and treatments [43]. Machine learning applications in epilepsy ranges from diagnosis of epilepsy [22, 44], PNES [45], and rare subtypes of epilepsy to the prevention of SUDEP and minimizing inter-observer variability of EEG interpretation.

Rajagopalan et al. showed that ML can diagnose temporal epilepsy by detecting microstate alterations than depending on ictal or interictal charges on repeated scalp EEG [44]. These, in turn, are affected by a multitude of factors such as medications, sleep deprivation, and inter-observer variability [46]. ML-detected alterations in microstate C shows a possible reflection of the inappropriate activation of alpha activity [47,48,49] with 76.1% accuracy, even in the absence of a visible interictal epileptiform (IED) discharges on EEG [44]. The EpiFinder algorithm used in a tertiary center was able to generate a differential of seizure types and epilepsy syndromes from other spells [22].

A pilot study using wristband sensors detected increased EDA, i.e., epidermal activity in epileptic seizures. The increase in EDA was proportionally higher in generalized tonic–clonic seizures (GTCS) versus complex partial seizures (CPS) [21]. Similar observations have been noted by Van Buren in 1958 [50], and elevated plasma catecholamines following GTCS support this assumption [51, 52]. They propose that this autonomic instability of sympathetic surge during seizures could play a role in SUDEP [21].

Psychogenic non-epileptic seizures (PNES) resemble epileptic seizures and consist of episodes of paroxysmal behavioral manifestations including a range of motor, sensory, and behavioral manifestations [53, 54]. 20% of the epilepsy patients referred to a tertiary center are eventually diagnosed with PNES using the gold standard video electroencephalography (vEEG); hence, there is a need to identify better, quicker, and affordable tests to reduce the significant chronic disability, lost wage hours, multiple hospitalizations, and associated risk of morbidity and mortality [53,54,55,56,57].

A group from Italy was able to identify PNES using ML with multivariate neuroimaging analysis and also secondarily locate brain regions within the limbic and the right inferior frontal cortex (IFC) [45]. IFC has also been implicated in other disorders such as compulsive–impulsive disorders, Tourette syndrome, and Parkinson's disease with levodopa-induced dyskinesia [58]. The EpiFinder algorithm used in a tertiary center was able to differentiate epileptic seizures from PNES [22].

Concussion

Concussion is another multidimensional problem with no validated criteria for diagnosis, leading to inter-examiner variability [59,60,61,62]. Its clinical presentation varies from cognitive to non-cognitive domains including sleep and balance [62,63,64,65,66]. Prior studies have focused on scans, symptoms, and cognitive testing [62, 65, 66] despite its varied symptomatology.

ML has not only been able to differentiate concussed and control subjects, and improve diagnosis based on individual data including conventional imaging, cognitive domains, eye movement, and clinical presentation [66,67,68,69,70,71,72,73,74], but has also made it possible to look into less understood and complex pathologies such as vestibular impairments [75,76,77,78,79], to better understand and identify varied phenotypes like cognitive problems, oculomotor dysfunction, affective disturbances, cervical spine disorders, headaches, and cardiovascular and vestibular anomalies [29, 80].

Dementia

Frontotemporal dementia (FTD) is a neurodegenerative disorder that accounts for 20% of young-onset dementia and has a high rate of misdiagnosis [81, 82]. This often leads to poor patient satisfaction and well-being, unnecessary laboratory tests, clinic visits, and imaging, and addition to the health-care costs [28, 83]. It has the worst prognosis and reduced life expectancy when compared to other dementias [12]. A study in the UK has shown that deep learning algorithms can reduce unnecessary investigations and improve costs and patient satisfaction through better clinical practice guidelines [28].

Movement disorders

The diagnosis of the two most common movement disorders, Parkinson’s disease (PD) and essential tremor (ET), is challenging and based primarily on clinical criteria [84]. Researchers are using novel methods to synchronize data into deep learning from tremor characteristics in varied environments, questionnaires addressing non-motor issues, and Archimedean spirals drawn to identify newer phenotypes [17]. Prior studies have used each of these characteristics individually to diagnose movement disorders [85].

Outcomes and limitations

As discussed above, AI has played an important role in identifying neurological disorders [13]. It has been transformational in converting the voluminous data collected into those of clinical relevance [13]. For all the benefits, however, there are huge limitations and unknown legal ramifications.

First, there is a lack of data standards and open data repositories in machine learning [7, 86]. For example, limitations in the integration of already available commercial medical devices such as Parkinson’s Kinetigraph TM into routine health care have been secondary to non-integration of motor and nonmotor characteristics, lacking open data storages and open programs [87]. Varghese et al. have overcome the limitations mentioned by not only integrating motor and non-motor phenomena, but also by using large data models with massive infrastructure like Medical-Data Models Portal (MDM), and combining affordable devices with customizable apps and SDS (smart device systems) that can be programmed by any Apple-based App developer [17]. This can help reach the level of data required to quicken the regulatory steps and rollout devices sooner with improved diagnostic accuracy [17, 84, 88] (Table 3).

Table 3 Benefits and barriers of artificial intelligence and machine learning
Full size table

Moreover, deep learning will never be 100% complete, especially with inter-examiner differences and varied environments [17]. Studies reveal that data entered by specialists improve performance in deep learning methods especially with pattern recognition [17].

In ML, there is some literature that the size of the samples should be a multiple of the number of input and output variables [19] nonetheless studies were affected by small sample size [17, 44, 45].

ML’s clinical application in confounding groups with similar neurological, psychological, or pathological presentation is limited [45]: for instance, ML’s application in distinguishing PNES not only from patients with epilepsy, but also with similar psychopathological presentation such as major depression [45], or ML’s ability to distinguish epilepsy from healthy controls [44] versus application in patients with panic disorder [89], schizophrenia [90], drug-induced, and memory deficit with similar alterations in microstate C [44].

AI or deep learning uses different reasoning methods to diagnose a disease. A systematic review by Arani et al. in multiple sclerosis showed different reasoning methods or a combination of methods, e.g., case based, rule based, model based, fuzzy logic, genetic algorithms, natural language processing, and neural networks were used by different computers and each method had its own capabilities and limitations [91]. The efficiency of each method varied, and, hence, affected its applicability in the diagnosis of rare and complicated disease such as multiple sclerosis [91]. Nonetheless, they can play a major role in helping patients and physicians with timely clinical diagnosis [91].

It is impossible to account for all potential positive or negative side effects even with the best algorithms [28]. Deep learning algorithms tend to avoid negative side effects and confounding factors like test results to achieve its goal, and thus in turn can affect patient safety and outcomes [28, 92].

Mobile apps are currently being used to monitor paroxysmal AF and long-term anticoagulation (39) and have shown to be effective tools in monitoring and screening of arrhythmias [18], but this has often led to a large number of false positives and in turn to expensive and unnecessary testing [18, 93]. This inadvertently raises the medico-legal aspect and logistics for governments and public health bodies with regard to systematic and opportunistic screening, cost-effectiveness, and management of newly identified patients with AF [18]; hence, a better understanding of local screening guidelines is required, e.g., European Heart Rhythm Association Consensus guidelines on screening AF.

Lastly, the ethical and legal ramifications are beyond the scope of this paper, but sustaining patient trust would be the key in promoting collaboration and implementation of AI [94].The black box scenario is that decisions made by AI-enabled computer-aided diagnosis (CAD) systems [95] cannot be explained. The legal ramifications of a misdiagnosis are unclear, especially with regard to whether the fault lies with the manufacturer or the physician [95,96,97]. We need to develop standard practices for evaluation of AI tools [98]. It is debatable whether AI will replace physicians, but AI will definitely be playing a greater role in integrating health care [94].

Future directions

Deep brain stimulation (DBS) is an effective surgical treatment option and improves quality of life in patients with tremor manifestation, including Parkinson’s (PD) and essential tremor (ET) [99,100,101,102]. Currently, determination of DBS leads is done by a neurologist, and is therefore affected by interpersonal variability. AI may help in this regard to make an objective evaluation, subject to fulfilling the regulatory requirements and getting medical clearance [17].

Using open data portals can be in the best shared interest to develop data standards and smart devices [86]. Some researchers have made the study models, frameworks, codes, and anonymized data samples open source, which makes it easier to reproduce in future studies [17]. Using devices that are affordable with customizable apps that are available to the masses increases their applicability [17]. For example, Kardiaband app by AliveCor is the first FDA-cleared smartwatch-based ECG reader [103].

Standard auditing and statistics only can test known hypotheses, but newer algorithms can test multiple hypotheses in a reasonable time frame and make a priori assumptions [28] beyond the scope of human capabilities. This has the potential to be applied across other diseases and specialties [28]. ML specifically is able to analyze large datasets at much quicker speeds and at higher accuracy to investigate unexplained phenomena [15]. ML algorithms like self-organizing map (SOM) can be extended to include non-vestibular parameters, including previous concussions, neuropsychological outcomes, and multiple other variables to validate studies in a more complicated context [29]. This also help to improve diagnostic criteria by identifying features among varied patient populations [16].

Conclusion

The current neurological care not only places a burden on the US economy by imposing higher overall costs, but also affects disability-adjusted life years. The ability of AI and ML to analyze medical data in disease prevention, diagnosis, patient monitoring, and development of new protocols will help clinicians to deal with voluminous data in a more accurate and efficient manner. AI has the ability to limit inter-observer variability, screen for asymptomatic atrial fibrillation, diagnose epilepsy, PNES, concussions, and movement disorders, and identify abnormal autonomic functions to prevent SUDEP. Though AI utilization is limited for a wide variety of reasons including physician’s reluctance to adopt novel technology, billing and reimbursement issues, pan-USA licensing problems, malpractice lawsuits, and the initial cost of technological establishment, it has the potential to serve as a powerful tool for neurological care.