Clinical Neurology A Primer

[Brook Mithani]

Thisis a clinical neurology book for the student, non neurologist, and those that teach them. The book covers neuroanatomy, history taking and examination and then proceeds to discuss the clinical features of common problems as well as some of the more common rare, neurological disorders, in a way that will demystify a part of medicine that students find complex and difficult to understand. The book is accompanied by a DVD explaining concepts, demonstrating techniques of performing the neurological examination and demonstration of abnormal neurological signs.

The first chapter is devoted to neuroanatomy from a clinical viewpoint. The concept of localising problems by likening the nervous system to a map grid with vertical meridians of longitude (the ascending sensory pathways and the descending motor pathway)and horizontal parallels of latitude (cortical signs, brainstem cranial nerves, nerve roots and peripheral nerves) of the nervous system is developed. Subsequent chapters take the reader through the neurological examination and the common neurological presentations from a symptom oriented approach. Chapter 4 contains a very simple method of understanding the brainstem, the "rule of 4". Chapter 6 discusses the approach after the history and examination are completed. The final chapter is an overview of how to approach information gathering and keeping up-to-date using the complex information streams available.

We reviewed foundational concepts in artificial intelligence (AI) and machine learning (ML) and discussed ways in which these methodologies may be employed to enhance progress in clinical trials and research, with particular attention to applications in the design, conduct, and interpretation of clinical trials for neurologic diseases. We discussed ways in which ML may help to accelerate the pace of subject recruitment, provide realistic simulation of medical interventions, and enhance remote trial administration via novel digital biomarkers and therapeutics. Lastly, we provide a brief overview of the technical, administrative, and regulatory challenges that must be addressed as ML achieves greater integration into clinical trial workflows.

The term artificial intelligence (AI) refers to the use of computational methods to enable machines to perform tasks such as perception, reasoning, learning, and decision-making. Advances in the technology sector are fueling the development of novel forms of AI, which are rapidly driving progress across diverse domains such as facial recognition, financial strategy, and self-driving vehicles [1, 2]. The field of medicine is no exception, with AI methods increasingly being applied in healthcare research, from the laboratory to the bedside. In clinical trials, particularly, automated methods similarly carry great promise to alleviate many of the considerable difficulties associated with planning, completing, and analyzing the results of large scale trials. The challenges associated with traditional trials, from recruiting participants across diverse populations to the selection of feasible and appropriate eligibility criteria, make these interventions an ideal area for the application of emerging data science techniques.

In this article, we reviewed machine learning (ML) as a means of achieving AI and improving the practice of clinical research. We provided a basic introduction to key ML concepts for clinicians, surveyed general areas of application for ML in clinical trials, and then demonstrated how ML is being used to foster innovation in clinical research for neurologic diseases, specifically. We concluded with a discussion of technical challenges to automation in trials, highlighting potential obstacles that must be overcome to sustain innovation in the field.

Efforts to standardize clinical care via advanced statistical models have their roots in the twentieth century [3, 4], when the advent of modern computers enabled researchers to begin simulating the process of differential diagnosis [3,4,5,6,7,8], recommending antibiotic regimens [9], and identifying medication effects [10]. Though these early initiatives fell short of making widespread impact [11], a number of factors have led to an unprecedented rate of progress in ML since the early 2010s.

Increased access to large quantities of electronic data (in medicine, most notably, publicly available datasets such as the UK Biobank [12] and the Cancer Genome Atlas [13]), advances in computer hardware (especially Graphics Processing Units [GPUs]), and the widespread availability of open source software [14] have created the necessary environment for AI to achieve significant gains. Furthermore, continued algorithmic developments have enabled machines to take on tasks of increasing complexity and nuance [15].

While the notion of learning implies some measure of human-like agency, medical ML algorithms depend on the transformation of patient-derived data into numerical formats that can be processed by computer systems. For instance, computed tomography (CT) scans can be understood as matrices of pixel intensities, and vital sign measurements may be translated into lists or vectors of discrete measurements. If an investigator can derive numerical quantities from a given data source, then the possibilities for which modalities can be used as input to an ML strategy are nearly limitless.

Lastly, the performance of medical ML models can be assessed according to a variety of metrics, depending on the specific use cases. In the case of diagnostic or prognostic classification tasks, it is often standard to report area under the receiver operating characteristic curve (AUROC), obtained by plotting true positive rate versus false positive rate at differing probability thresholds when comparing predictions versus observation [39]. Area under the precision-recall curve (AUPR) (obtained from plotting positive predictive value versus sensitivity) may also be reported, as AUROC may overestimate performance in the case of highly imbalanced datasets [40]. A variety of specialized metrics for tasks such as segmentation (e.g., dice coefficient and intersection-over-union) [41], image generation (e.g., structural similarity) [42], and other tasks may also be deployed depending on the use case. Conversely, in regression for continuous quantities, standard metrics such as the mean squared error (MSE) between predicted and observed values may also be used [43]. Regardless of the specific measure employed, however, it is also imperative that ML models be judged in terms of traditional criteria (e.g., sensitivity, specificity, accuracy) in order to fully contextualize their impact on patient care prior to deployment. An overview of essential ML terminology along with definitions is provided in Table 1. Examples of widely used ML algorithms are illustrated in Fig. 1 and further elaborated in Table 2.

Despite their successes, modern clinical trials remain difficult for research teams to bring to completion. Remarkably, unsuccessful trials remain the norm rather than exception due to myriad difficulties in identifying, enrolling, and providing treatment to patients within RCTs. Indeed, it has been estimated that only 12% of drug development programs achieve clinical trial success from phase 1 to launch [59]. While lack of clinical efficacy makes up a large component of the failures, many clinical studies fall short of recruitment goals and timelines due to factors such as low patient participation in clinical research and overly stringent inclusion criteria [60].

Moreover, the dramatic increase in the availability of electronic health records (EHR) due to advances in information technology [62] has complicated the task of examining available data for identifying and pre-screening potential research participants. Ostensibly, the growth of health records has created both challenges and opportunities [63]. The International Classification of Disease (ICD) diagnostic codes used worldwide for clinical billing, for instance, could potentially be used to identify patients who have the condition of interest. However, diagnostic codes may also be misapplied by treating clinicians [64, 65], reflecting outdated or suspected but unconfirmed diagnoses. This inconsistency within EHRs not only complicates efforts for maintaining an accurate clinical record but also affects the ability of research staff to leverage large databases to accurately pre-screen for clinical trials. Automated methods for maintaining an accurate medical history could be a particularly useful innovation.

Lastly, in an age of increasing awareness of healthcare inequality, ML methods for patient recruitment may be applied to alleviate racial disparities in clinical trials. Notably, it has been estimated that nearly 90% of participants in these studies are White [75], while historical surveys of clinical trials show that they are poorly representative of women, ethnic minorities, and patients outside of relatively wealthy regions such as North America or Western Europe [76, 77]. There is little doubt that drug and medical device development poses the risk of further alienating disadvantaged patient populations when ML-based methods used to validate them in clinical trials rely on data from non-representative groups [78, 79]. The generalizability gap, however, may in part be alleviated by automated methods for improving enrollment of historically underserved groups. Zhang and colleagues, for instance, have demonstrated the usage of ML classifiers to explicitly match pregnant women and persons living with HIV to oncology trials from ClinicalTrials.gov [80]. Health systems may also use enhanced screening capacity for trial eligibility to match patients from excluded groups to ongoing studies, either by NLP methods that explicitly take into account patient identities or from the types of data-driven eligibility expansions proposed by Liu and colleagues [81, 82]. Electronic phenotyping of disease characteristics rather than demographic factors may also identify which patients are most appropriate for enrollment on the basis of their physical health, though certain clinical phenotypes (e.g., poor pulmonary function and high BMI) may retain confounding relationships with race, ethnicity, class, and gender [83]. To enhance diversity in clinical trials, a promising strategy is to use ML to identify clinical sites that may benefit from focused resources aimed at training and recruiting investigative site personnel from underrepresented minority groups. These efforts can lead to a greater representation of diverse participants in clinical trials, underscoring the importance of prioritizing such initiatives.

3a8082e126