Will Artificial Intelligence Outperform the Clinical Neurologist in the Near Future? Yes

In the article “Predicting the Future—Big Data, Machine Learning, and Clinical Medicine,” published in the New England Journal of Medicine, it is speculated that machine learning (ML) will soon displace much of the work of radiologists and anatomical pathologists. ¹ For example, algorithms can replace the need for a second radiologist reading of mammograms, and recent studies have shown that machine accuracy is identical or exceeds that of humans in these fields. ² , ³ So is the same happening in neurology? Are we going to be outperformed by artificial intelligence (AI)?

As a neurologist with a keen interest in technologies, I hope not, but I do think AI is going to change what we do. Therefore, neurologists need to stay aware of these innovations and be prepared to embrace the new reality they will bring. To demonstrate, I draw your attention to the journey of a patient with young‐onset Parkinson's disease (YOPD). This real case shows where, in clinical practice, AI can support neurologists as assistant, monitor, coach, and teammate. ⁴

Clinical Vignette Part 1

At the age of 35, the patient began to notice that her left shoe frequently came off her foot during fast walking. The symptom continued for years, associated with mild left foot cramp and shoulder pain in the early morning. She was initially treated for tendinitis by several physicians before being referred to a neurologist, where the diagnosis of Parkinson's disease (PD) was made based on the examination of asymmetric parkinsonism. She initially denied the diagnosis as it was based solely on the examination without objective evidence. Her belief was that PD is a disease of an old age.

AI in the Diagnosis of PD

Unlike certain cases in radiology or pathology, where a diagnosis focuses on benign or malignant pathologies, the diagnosis of PD is not a binary judgment. It involves a process of clinical cognition, beginning with aggregation of findings, followed by selection of a pivot, generation of a cause list, and pruning the cause list before arriving at a list of possible diagnoses. ⁵ As more information becomes available from guided investigations, possibilities are narrowed down, and a final diagnosis emerges. While little is known about the cognitive processes that are employed in the solution of clinical problems, one study back in 1976, using a computer as a laboratory for the study of clinical cognition, revealed that the process of refinement, testing, and revision resembled the goal‐directed programming with pattern matching seen in computer science. ⁶ In addition to disease knowledge, the program used small problem‐solving strategies, similar to human “common sense,” to pursue interrelated information and deliver a final diagnosis. This early publication shows that the idea of using a computer to aid or replace clinicians is not new and simulation of the cognitive process can be demonstrated in a form of a funnel, its wide start representing the reasoning process just before the point of the first contact (1), with the shrinking diameter reflecting the narrowing down of possibilities with progressive acquisition of new information (2) until the narrow end, the final conclusion (3) (Fig. 1). ⁷ , ⁸ The ability of clinicians to go from point A to point B does not probably depend on only existing knowledge but also on individual clinical experience. Medical students would apply their recently acquired knowledge on the diagnostic criteria of parkinsonism to the case, similar to rule‐based expert systems, which rely on prefed static data. Conversely, seasoned neurologists would be able to narrow down the causes of left foot cramp and shoulder pain (1) and interpret them as supporting features of parkinsonism (2) by using the most effective mental skill—that of comparing patterns, which goes beyond that inherent in any formal set of diagnostic criteria. ⁹ This process is akin to ML as it learns rules from data. The process from A to B not only requires a neurologist's prior knowledge but also his or her intuition and instincts to judge what information is relevant, similar to “common sense,” and as such is unlikely to be surpassed by AI. ⁷ If we chose to implement AI to achieve this level of clinical reasoning, the programmer would have to develop a means of representing the boundless possibilities at this point. To the best of my knowledge, such advances of knowledge engineering, which aim to solve problems that usually require a high level of human expertise, still remains to be accomplished. ¹⁰

FIG 1.

Funnel plot of diagnostic clinical reasoning in neurology. See text for discussion. AI, artificial intelligence.

Despite these limitations, AI still has an enormous potential in the management of PD, especially when it is operating in so‐called “microworlds,” that is, within extremely small task domains that are near point B when the possibilities have become limited. At this point, AI can serve as assistant and teammate to capture and analyze the full complexity of clinical and multidimensional imaging data that is beyond traditional mathematical analysis. Indeed, recent studies have provided promising results demonstrating that ML improves classification performance of dopamine transporter single photon emission computed tomography images for the differentiation of PD from other parkinsonian disorders, potentially shortening the diagnostic delay in our case of YOPD. ¹¹ , ¹² This application could be implemented in underserved regions assisting a remote radiologist to interpret advanced neuroimages, leading to an earlier and more accurate diagnosis of PD, especially in the early stage when characteristic features to confirm or refute the diagnosis are quite subtle and neither sensitive enough to be observed by clinical examination (eg, subtle red flags) nor specific enough for clinical interpretation (eg, a less defined dopaminergic response). ¹³

Clinical Vignette Part 2

The patient responded well to dopaminergic medications and worked full‐time as an accountant so she was very keen to know the long‐term prognosis. Two years following treatment with dopaminergic medications, she developed complex fluctuations with several episodes of disabling “off” periods and nocturnal hypokinesia. The patient was being considered for deep brain stimulation (DBS) and wanted to know her personal risks. She lived far from her neurologist and became very concerned if she would not be able to manage this device effectively.

AI in the Management of PD

ML will dramatically improve the ability of healthcare professionals to establish a prognosis of PD. Traditional prognostic models are restricted to only a handful of clinical variables, relying on cardinal motor features to demonstrate slower disease progression on tremor‐predominant rather than akinetic‐rigid or postural instability subtypes. AI could assist neurologists in the extrapolation of data from electronic health records, claim databases, genetic and transcriptomic data, or various objective measures allowing models to use thousands of rich predictor variables. ML models that combine genetic data with demographic, clinical, and neuroimaging information have achieved significant refinement in PD diagnosis and disease phenotype prediction as well as PD subtype identification. In a longitudinal analysis of patient data from different cohorts, genetic information has proved to be more indicative of motor progression (genetic variation, rs17710829 and rs9298897) and global cognitive impairment (GBA mutation) than other clinical features. ¹⁴ , ¹⁵ Among clinical variables, nonmotor symptoms (NMS) were found to be the main determinants of rapid symptoms progression within the first 2‐year and 4‐year follow‐up periods, with autonomic dysfunction, mood impairment, anxiety, rapid eye movement sleep behavior disorders, and cognitive decline as clinical predictors at baseline evaluation. ¹⁶

The rapid growth of wearable sensors also offers new insights into the nature and characteristics of various motor and NMS in PD. ¹⁷ The concept of digital phenotyping has been applied in PD where the moment‐by‐moment quantification of the individual‐level patient phenotype with a personal digital device is now possible. ¹⁸ These capabilities offer several advantages for individualized and remote assessment of patients with PD in their own environment that directly determines a patient's performance, not capacity to recall as with most rating scales, in their daily tasks, and extends the opportunity for remote assessment during the current pandemic situation or in the underserved regions where face‐to‐face evaluations are not possible. ¹⁹ , ²⁰ Recent evidence, which could be applied to our example case, suggested that ML can analyze wearable data, providing feedback information on motor performances to neurologists for the adjustment of dopaminergic medications resulting in improved clinical outcomes. ²¹ Further to this, the use of mobile health technologies equipped with AI could accelerate the transition of a single objective domain assessment (eg, tremor or dyskinesia) to a set of activities that are part of patient‐centered digital outcome measures with various applications being implemented into real‐world assessments and novel therapeutic evaluations. ²²

One of the key challenges for neurologists is to properly identify suitable candidates for DBS or other device‐aided therapies who will receive measurable clinical benefits with minimal surgical and stimulation‐related risks. ML has been applied in a retrospective cohort of >500 patients with DBS and identified presurgical features that were associated with a greater chance of DBS complications, including age older than 75 years, body mass index over 25, diabetes mellitus, and history of smoking among others. ²³ With the recent surge of the COVID‐19 pandemic, most neurologists have embraced telemedicine as a method of reducing the risk of contagion, but are also seeing other benefits of improved care, convenience, comfort, and confidentiality for patients with PD. ²⁴ In addition to the standard assessment during a telemedicine visit, AI can be a neurologist's teammate, providing accurate evaluations of patients with PD remotely based on continuous signals from wearables that are transformed into a sequence of “syllables” that are disorganized in patients with PD and correlated with disease severity. ²⁵

Conclusion

The case vignettes highlight the various incidences where AI can be applied during the natural history of PD. It seems clear that AI and neurologists are most potent when they cooperate. It should never be AI versus neurologists since AI‐based applications serve to help us help patients. However, in some areas such as advanced neuroimaging or technology‐based monitoring, AI may advance clinical reasoning in the interpretation. The benefits of AI have recently been demonstrated in other fields of movement disorders beyond PD as shown in cervical dystonia when deep learning was used to classify it into 3 types and automatically segmenting individual muscles for targeted injections. ²⁶ No matter whether it is AI, robotics, or augmented or virtual reality, we should accept that these advances are going to have a massive influence on the way we work. I believe that the following quote from Charles Darwin is still applicable: “It is not the strongest of the species that survive, nor the most intelligent, but the one most responsive to change.” Modern AI research only began in the 1950s, whereas our brain, which is more powerful than any machine in existence, has been evolving for hundreds of millions of years. As a neurologist in this generation, I am optimistic that we will continue to evolve and embrace this new reality.

Disclosures

Ethical Compliance Statement

The author confirms that neither patient consent nor the ethical approval of an institutional review board was required for this work. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this work is consistent with those guidelines.

Funding Sources and Conflict of Interest

No specific funding was received for this work. The author declares that there are no conflicts of interest relevant to this work.

Financial Disclosures for the Previous 12 Months

Roongroj Bhidayasiri is supported by the Senior Research Scholar Grant (RTA6280016) of the Thailand Science Research and Innovation, an International Research Network Grant of the Thailand Research Fund (IRN59W0005), a Chulalongkorn Academic Advancement Fund into Its 2nd Century Project of Chulalongkorn University, and a Centre of Excellence grant of Chulalongkorn University (GCE 6100930004‐1), Bangkok, Thailand. He receives a salary from Chulalongkorn University and stipend from the Royal Society of Thailand; has received consultancy and/or honoraria/lecture fees from Abbott, Boehringer‐Ingelheim, Britannia, Ipsen, Novartis, Teva‐Lundbeck, Takeda, and Otsuka pharmaceuticals; he holds patents for a laser‐guided walking stick, portable tremor device, nocturnal monitoring, and electronic Parkinson's disease symptom diary as well as copyright on a Parkinson's mascot, dopamine lyrics, and teaching video clips for common nocturnal and gastrointestinal symptoms for Parkinson's disease.