Big data, machine learning and artificial intelligence: a neurologist's guide.

Modern clinical practice requires the integration and interpretation of ever-expanding volumes of clinical data. There is, therefore, an imperative to develop efficient ways to process and understand these large amounts of data. Neurologists work to understand the function of biological neural networks, but artificial neural networks and other forms of machine learning algorithm are likely to be increasingly encountered in clinical practice. As their use increases, clinicians will need to understand the basic principles and common types of algorithm. We aim to provide a coherent introduction to this jargon-heavy subject and equip neurologists with the tools to understand, critically appraise and apply insights from this burgeoning field.

Retrosplenial Cortex Indexes Stability beyond the Spatial Domain.

Retrosplenial cortex (RSC) is highly responsive to landmarks in the environment that remain fixed in a permanent location, and this has been linked with its known involvement in scene and spatial processing. However, it is unclear whether RSC representations of permanence are a purely spatial phenomenon or whether they extend into behavioral and conceptual domains. To test this, during functional MRI scanning, we had people (males and females) read three different types of sentences that described either something permanent or transient. The first two sentence types were imageable, with a focus either on a spatial landmark or on an action. The third type of sentence involved non-imageable abstract concepts. We found that, in addition to being more active for sentences describing landmarks with a permanent location in space, RSC was also significantly engaged by sentences describing stable and consistent behaviors or actions, as long as they were rooted within a concrete imageable setting. RSC was not responsive to abstract concepts, even those that embodied the notion of stability. Similarly, it was not engaged by imageable sentences with transient contents. In contrast, parahippocampal cortex was more engaged by imageable sentences describing landmarks, whereas the hippocampus was active for all imageable sentences. In addition, for imageable sentences describing permanence, there was bidirectional functional coupling between RSC and these medial temporal lobe structures. It appears, therefore, that RSC-mediated permanence representations could be helpful for more than spatially mapping environments and may also provide information about the reliability of events occurring within them.SIGNIFICANCE STATEMENT The retrosplenial cortex (RSC) is known to process information about landmarks in the environment that have a fixed, permanent location. Here we tested whether this permanence response was apparent beyond the spatial domain, which could have implications for understanding the role of the RSC more widely across cognition. We found that the RSC was engaged not only by permanent landmarks but also by stable and consistent actions. It was not responsive to transient landmarks or actions or to abstract concepts, even those that embodied the notion of stability. We conclude that the RSC might do more than help to map spatial environments, by possibly also providing information about the reliability of events occurring within them.

Dissociating Landmark Stability from Orienting Value Using Functional Magnetic Resonance Imaging.

Retrosplenial cortex (RSC) plays a role in using environmental landmarks to help orientate oneself in space. It has also been consistently implicated in processing landmarks that remain fixed in a permanent location. However, it is not clear whether the RSC represents the permanent landmarks themselves or instead the orienting relevance of these landmarks. In previous functional magnetic resonance imaging (fMRI) studies, these features have been conflated-stable landmarks were always useful for orienting. Here, we dissociated these two key landmark attributes to investigate which one best reflects the function of the RSC. Before scanning, participants learned the features of novel landmarks about which they had no prior knowledge. During fMRI scanning, we found that the RSC was more engaged when people viewed permanent compared with transient landmarks and was not responsive to the orienting relevance of landmarks. Activity in RSC was also related to the amount of landmark permanence information a person had acquired and, as knowledge increased, the more the RSC drove responses in the anterior thalamus while viewing permanent landmarks. In contrast, the angular gyrus and the hippocampus were engaged by the orienting relevance of landmarks, but not their permanence, with the hippocampus also sensitive to the distance between relevant landmarks and target locations. We conclude that the coding of permanent landmarks in RSC may drive processing in regions like anterior thalamus, with possible implications for the efficacy of functions such as navigation.

Improving estimation of Parkinson's disease risk-the enhanced PREDICT-PD algorithm.

We previously reported a basic algorithm to identify the risk of Parkinson's disease (PD) using published data on risk factors and prodromal features. Using this algorithm, the PREDICT-PD study identified individuals at increased risk of PD and used tapping speed, hyposmia and REM sleep behaviour disorder (RBD) as "intermediate" markers of prodromal PD in the absence of sufficient incident cases. We have now developed and tested an enhanced algorithm which incorporates the intermediate markers into the risk model. Risk estimates were compared using the enhanced and the basic algorithm in members of the PREDICT-PD pilot cohort. The enhanced PREDICT-PD algorithm yielded a much greater range of risk estimates than the basic algorithm (93-609-fold difference between the 10th and 90th centiles vs 10-13-fold respectively). There was a greater increase in the risk of PD with increasing risk scores for the enhanced algorithm than for the basic algorithm (hazard ratios per one standard deviation increase in log risk of 2.75 [95% CI 1.68-4.50; p < 0.001] versus 1.47 [95% CI 0.86-2.51; p = 0.16] respectively). Estimates from the enhanced algorithm also correlated more closely with subclinical striatal DaT-SPECT dopamine depletion (R2 = 0.164, p = 0.005 vs R2 = 0.043, p = 0.17). Incorporating the previous intermediate markers of prodromal PD and using likelihood ratios improved the accuracy of the PREDICT-PD prediction algorithm.

Optimising classification of Parkinson's disease based on motor, olfactory, neuropsychiatric and sleep features.

Olfactory loss, motor impairment, anxiety/depression, and REM-sleep behaviour disorder (RBD) are prodromal Parkinson's disease (PD) features. PD risk prediction models typically dichotomize test results and apply likelihood ratios (LRs) to scores above and below cut-offs. We investigate whether LRs for specific test values could enhance classification between PD and controls. PD patient data on smell (UPSIT), possible RBD (RBD Screening Questionnaire), and anxiety/depression (LADS) were taken from the Tracking Parkinson's study (n = 1046). For motor impairment (BRAIN test) in PD cases, published data were supplemented (n = 87). Control data (HADS for anxiety/depression) were taken from the PREDICT-PD pilot study (n = 1314). UPSIT, RBDSQ, and anxiety/depression data were analysed using logistic regression to determine which items were associated with PD. Gaussian distributions were fitted to BRAIN test scores. LRs were calculated from logistic regression models or score distributions. False-positive rates (FPRs) for specified detection rates (DRs) were calculated. Sixteen odours were associated with PD; LRs for this set ranged from 0.005 to 5511. Six RBDSQ and seven anxiety/depression questions were associated with PD; LRs ranged from 0.35 to 69 and from 0.002 to 402, respectively. BRAIN test LRs ranged from 0.16 to 1311. For a 70% DR, the FPR was 2.4% for the 16 odours, 4.6% for anxiety/depression, 16.0% for the BRAIN test, and 20.0% for the RBDSQ. Specific selections of (prodromal) PD marker features rather than dichotomized marker test results optimize PD classification. Such optimized classification models could improve the ability of algorithms to detect prodromal PD; however, prospective studies are needed to investigate their value for PD-prediction models.

Testing Shortened Versions of Smell Tests to Screen for Hyposmia in Parkinson's Disease.

BACKGROUND: Hyposmia is an early feature in neurodegenerative diseases, most notably Parkinson's disease (PD). Using abbreviated smell tests could provide a cost-effective means for large-scale hyposmia screening. It is unclear whether short smell tests can effectively detect hyposmia in patient populations. OBJECTIVES: To test the ability of short smell combinations to "prescreen" for probable hyposmia in people with PD and target administration of more extensive tests, such as the University of Pennsylvania Smell Identification Test. METHODS: We assessed the screening performance of a short 4-smell combination previously derived from use of the 40-item University of Pennsylvania Smell Identification Test in healthy older people and its ability to detect hyposmia in a large cohort of PD patients. RESULTS: The novel 4-smell combination included menthol, clove, onion, and orange and had a sensitivity of 87.1% (95% confidence interval, 84.9%-89.2%) and specificity of 69.7% (63.3%-75.5%) for detecting hyposmia in patients with PD. A different (also novel) 4-item combination developed using a data-driven approach in PD patients only achieved 81.3% (78.2%-84.4%) sensitivity for equivalent specificity. CONCLUSIONS: A short 4-smell combination derived from a healthy population demonstrated high sensitivity to detect those with hyposmia and PD.

Screening performance of abbreviated versions of the UPSIT smell test.

BACKGROUND: Hyposmia can develop with age and in neurodegenerative conditions, including Parkinson's disease (PD). The University of Pennsylvania Smell Identification Test (UPSIT) is a 40-item smell test widely used for assessing hyposmia. However, in a number of situations, such as identifying hyposmic individuals in large populations, shorter tests are preferable. METHODS: We assessed the ability of shorter UPSIT subsets to detect hyposmia in 891 healthy participants from the PREDICT-PD study. Shorter subsets included Versions A and B of the 4-item Pocket Smell Test (PST) and 12-item Brief Smell Identification Test (BSIT). Using a data-driven approach, we evaluated screening performances of 23,231,378 combinations of 1-7 smell items from the full UPSIT to derive "winning" subsets, and validated findings separately in another 191 healthy individuals. We then compared discriminatory UPSIT smells between PREDICT-PD participants and 40 PD patients, and assessed the performance of "winning" subsets containing discriminatory smells in PD patients. RESULTS: PST Versions A and B achieved sensitivity/specificity of 76.8%/64.9% and 86.6%/45.9%, respectively, while BSIT Versions A and B achieved 83.1%/79.5% and 96.5%/51.8%. From the data-driven analysis, 2 "winning" 7-item subsets surpassed the screening performance of 12-item BSITs (validation sensitivity/specificity of 88.2%/85.4% and 100%/53.5%), while a "winning" 4-item subset had higher sensitivity than PST-A, -B, and even BSIT-A (validation sensitivity 91.2%). Interestingly, several discriminatory smells featured within "winning" subsets, and demonstrated high-screening performances for identifying hyposmic PD patients. CONCLUSION: Using abbreviated smell tests could provide a cost-effective means of large-scale hyposmia screening, allowing more targeted UPSIT administration in general and PD-related settings.

Efficacy of navigation may be influenced by retrosplenial cortex-mediated learning of landmark stability.

Human beings differ considerably in their ability to orient and navigate within the environment, but it has been difficult to determine specific causes of these individual differences. Permanent, stable landmarks are thought to be crucial for building a mental representation of an environment. Poor, compared to good, navigators have been shown to have difficulty identifying permanent landmarks, with a concomitant reduction in functional MRI (fMRI) activity in the retrosplenial cortex. However, a clear association between navigation ability and the learning of permanent landmarks has not been established. Here we tested for such a link. We had participants learn a virtual reality environment by repeatedly moving through it during fMRI scanning. The environment contained landmarks of which participants had no prior experience, some of which remained fixed in their locations while others changed position each time they were seen. After the fMRI learning phase, we divided participants into good and poor navigators based on their ability to find their way in the environment. The groups were closely matched on a range of cognitive and structural brain measures. Examination of the learning phase during scanning revealed that, while good and poor navigators learned to recognise the environment's landmarks at a similar rate, poor navigators were impaired at registering whether landmarks were stable or transient, and this was associated with reduced engagement of the retrosplenial cortex. Moreover, a mediation analysis showed that there was a significant effect of landmark permanence learning on navigation performance mediated through retrosplenial cortex activity. We conclude that a diminished ability to process landmark permanence may be a contributory factor to sub-optimal navigation, and could be related to the level of retrosplenial cortex engagement.

A central role for the retrosplenial cortex in de novo environmental learning.

With experience we become accustomed to the types of environments that we normally encounter as we navigate in the world. But how does this fundamental knowledge develop in the first place and what brain regions are involved? To examine de novo environmental learning, we created an 'alien' virtual reality world populated with landmarks of which participants had no prior experience. They learned about this environment by moving within it during functional MRI (fMRI) scanning while we tracked their evolving knowledge. Retrosplenial cortex (RSC) played a central and highly selective role by representing only the most stable, permanent features in this world. Subsequently, increased coupling was noted between RSC and hippocampus, with hippocampus then expressing knowledge of permanent landmark locations and overall environmental layout. Studying how environmental representations emerge from scratch provided a new window into the information processing underpinning the brain's navigation system, highlighting the key influence of the RSC.

Assessing the mechanism of response in the retrosplenial cortex of good and poor navigators.

The retrosplenial cortex (RSC) is consistently engaged by a range of tasks that examine episodic memory, imagining the future, spatial navigation, and scene processing. Despite this, an account of its exact contribution to these cognitive functions remains elusive. Here, using functional MRI (fMRI) and multi-voxel pattern analysis (MVPA) we found that the RSC coded for the specific number of permanent outdoor items that were in view, that is, items which are fixed and never change their location. Moreover, this effect was selective, and was not apparent for other item features such as size and visual salience. This detailed detection of the number of permanent items in view was echoed in the parahippocampal cortex (PHC), although the two brain structures diverged when participants were divided into good and poor navigators. There was no difference in the responsivity of the PHC between the two groups, while significantly better decoding of the number of permanent items in view was possible from patterns of activity in the RSC of good compared to poor navigators. Within good navigators, the RSC also facilitated significantly better prediction of item permanence than the PHC. Overall, these findings suggest that the RSC in particular is concerned with coding the presence of every permanent item that is in view. This mechanism may represent a key building block for spatial and scene representations that are central to episodic memories and imagining the future, and could also be a prerequisite for successful navigation.

Retrosplenial cortex codes for permanent landmarks.

Landmarks are critical components of our internal representation of the environment, yet their specific properties are rarely studied, and little is known about how they are processed in the brain. Here we characterised a large set of landmarks along a range of features that included size, visual salience, navigational utility, and permanence. When human participants viewed images of these single landmarks during functional magnetic resonance imaging (fMRI), parahippocampal cortex (PHC) and retrosplenial cortex (RSC) were both engaged by landmark features, but in different ways. PHC responded to a range of landmark attributes, while RSC was engaged by only the most permanent landmarks. Furthermore, when participants were divided into good and poor navigators, the latter were significantly less reliable at identifying the most permanent landmarks, and had reduced responses in RSC and anterodorsal thalamus when viewing such landmarks. The RSC has been widely implicated in navigation but its precise role remains uncertain. Our findings suggest that a primary function of the RSC may be to process the most stable features in an environment, and this could be a prerequisite for successful navigation.