Objective This article methodically reviews the literature on deep learning (DL) for natural language processing (NLP) in the clinical domain, providing quantitative analysis to answer 3 research questions concerning methods, scope, and context of current research. Materials and Methods We searched MEDLINE, EMBASE, Scopus, the Association for Computing Machinery Digital Library, and the Association for Computational Linguistics Anthology for articles using DL-based approaches to NLP problems in electronic health records. After screening 1,737 articles, we collected data on 25 variables across 212 papers. Results DL in clinical NLP publications more than doubled each year, through 2018. Recurrent neural networks (60.8%) and word2vec embeddings (74.1%) were the most popular methods; the information extraction tasks of text classification, named entity recognition, and relation extraction were dominant (89.2%). However, there was a “long tail” of other methods and specific tasks. Most contributions were methodological variants or applications, but 20.8% were new methods of some kind. The earliest adopters were in the NLP community, but the medical informatics community was the most prolific. Discussion Our analysis shows growing acceptance of deep learning as a baseline for NLP research, and of DL-based NLP in the medical community. A number of common associations were substantiated (eg, the preference of recurrent neural networks for sequence-labeling named entity recognition), while others were surprisingly nuanced (eg, the scarcity of French language clinical NLP with deep learning). Conclusion Deep learning has not yet fully penetrated clinical NLP and is growing rapidly. This review highlighted both the popular and unique trends in this active field.

The presence of large number of false lesion classification on segmented brain MR images is a major problem in the accurate determination of lesion volumes in multiple sclerosis (MS) brains. In order to minimize the false lesion classifications, a strategy that combines parametric and nonparametric techniques is developed and implemented. This approach uses the information from the proton density (PD)-and T2-weighted and fluid attenuation inversion recovery (FLAIR) images. This strategy involves CSF and lesion classification using the Parzen window classifier. Image processing, morphological operations and ratio maps of PD and T2 weighted images are used for minimizing false positives. Contextual information is exploited for minimizing the false negative lesion classifications using hidden Markov random field -expectation maximization (HMRF-EM) algorithm. Lesions are delineated using fuzzy connectivity. The performance of this algorithm is quantitatively evaluated on 23 MS patients. Similarity index, percentages of over, under and correctestimations of lesions are computed by spatially comparing the results of present procedure with expert manual segmentation. The automated processing scheme detected 80% of the manually segmented lesions in the case of low-lesion load and 93% of the lesions in those cases with high lesion load.

Background Accurate classification of MS lesions in the brain cortex may be important in understanding their impact on cognitive impairment. Improved accuracy in identification/classification of cortical lesions was demonstrated in a study combining two MRI sequences: double inversion recovery (DIR) and T1-weighted phase-sensitive inversion recovery (PSIR). Objective To evaluate the role of intracortical lesions (IC) in MS related cognitive impairment (CI) and compare it to the role of mixed (MX), juxtacortical (JX), the sum of IC + MX and with total lesions as detected on DIR/PSIR images. Correlations between CI and brain atrophy, disease severity and disease duration were also sought. Methods 39 patients underwent extensive neuropsychological testing and were classified into: normal and impaired. Images were obtained on a 3T scanner and cortical lesions were assessed blind to the cognitive status of the subjects. Results 238 cortical lesions were identified (130 IC, 108 MX) in 82% of the patients, 39 JX lesions were also identified. Correlations between CI and MX lesions alone (p=0.010) and with the sum of IC + MX lesions (p=0.030) were found. A correlation between severity of CI and EDSS was also seen (p=0.009). Conclusion Cortical lesions play an important role in CI. However our results suggest that lesions that remain contained within the cortical ribbon do not play a more important role than ones extending into the adjacent white matter; furthermore the size of the cortical lesion, and not the tissue specific location, may better explain their correlation with CI.

Tensor based morphometry (TBM) was applied to determine the atrophy of deep gray matter (DGM) structures in 88 relapsing multiple sclerosis (MS) patients. For group analysis of atrophy, an unbiased atlas was constructed from 20 normal brains. The MS brain images were co-registered with the unbiased atlas using a symmetric inverse consistent nonlinear registration. These studies demonstrate significant atrophy of thalamus, caudate nucleus, and putamen even at a modest clinical disability, as assessed by the expanded disability status score (EDSS). A significant correlation between atrophy and EDSS was observed for different DGM structures: (thalamus: r = −0.51, p = 3.85×10 −7 ; caudate nucleus: r = −0.43, p = 2.35×10 −5 ; putamen: r = −0.36, p = 6.12×10 −6 ). Atrophy of these structures also correlated with 1) T2 hyperintense lesion volumes (thalamus: r = −0.56, p = 9.96×10 −9 ; caudate nucleus: r = −0.31, p = 3.10×10 −3 ; putamen: r = −0.50, p = 6.06×10 −7 ), 2) T1 hypointense lesion volumes (thalamus: r = −0.61, p = 2.29×10 −10 ; caudate nucleus: r = −0.35, p = 9.51×10 −4 ; putamen: r = −0.43, p = 3.51×10 −5 ), and 3) normalized CSF volume (thalamus: r = −0.66, p = 3.55×10 −12 ; caudate nucleus: r = −0.52, p = 2.31×10 −7 , and putamen: r = −0.66, p = 2.13×10 −12 ). More severe atrophy was observed mainly in thalamus at higher EDSS. These studies appear to suggest a link between the white matter damage and DGM atrophy in MS.

A comprehensive analysis of the global and regional values of cortical thickness based on 3D magnetic resonance images was performed on 250 relapsing remitting multiple sclerosis (MS) patients who participated in a multi-center, randomized, phase III clinical trial (the CombiRx Trial) and 125 normal controls. The MS cohort was characterized by relatively low clinical disability and short disease duration. An automatic pipeline was developed for identifying images with poor quality and artifacts. The global and regional cortical thicknesses were determined using FreeSurfer software. Our results indicate significant cortical thinning in multiple regions in the MS patient cohort relative to the controls. Both global cortical thinning and regional cortical thinning were more prominent in the left hemisphere relative to the right hemisphere. Modest correlation was observed between cortical thickness and clinical measures that included the extended disability status scale and disease duration. Modest correlation was also observed between cortical thickness and T1-hypointense and T2-hyperintense lesions. These correlations were very similar at 1.5 T and 3 T field strengths. A much weaker inverse correlation between cortical thickness and age was observed among the MS subjects compared to normal controls. This age-dependent correlation was also stronger in males than in females. The values of cortical thickness were very similar at 1.5 T and 3 T field strengths. However, the age-dependent changes in both global and regional cortical thicknesses were observed to be stronger at 3 T relative to 1.5 T.