Influence of coronary stenosis location on diagnostic performance of machine learning-based fractional flow reserve from CT angiography

BackgroundCompared with invasive fractional flow reserve (FFR), coronary CT angiography (cCTA) is limited in detecting hemodynamically relevant lesions. cCTA-based FFR (CT-FFR) is an approach to overcome this insufficiency by use of computational fluid dynamics. Applying recent innovations in computer science, a machine learning (ML) method for CT-FFR derivation was introduced and showed improved diagnostic performance compared to cCTA alone. We sought to investigate the influence of stenosis location in the coronary artery system on the performance of ML-CT-FFR in a large, multicenter cohort.MethodsThree hundred and thirty patients (75.2% male, median age 63 years) with 502 coronary artery stenoses were included in this substudy of the MACHINE (Machine Learning Based CT Angiography Derived FFR: A Multi-Center Registry) registry. Correlation of ML-CT-FFR with the invasive reference standard FFR was assessed and pooled diagnostic performance of ML-CT-FFR and cCTA was determined separately for the following stenosis locations: RCA, LAD, LCX, proximal, middle, and distal vessel segments.ResultsML-CT-FFR correlated well with invasive FFR across the different stenosis locations. Per-lesion analysis revealed improved diagnostic accuracy of ML-CT-FFR compared with conventional cCTA for stenoses in the RCA (71.8% [95% confidence interval, 63.0%–79.5%] vs. 54.8% [45.7%–63.8%]), LAD (79.3 [73.9–84.0] vs. 59.6 [53.5–65.6]), LCX (84.1 [76.0–90.3] vs. 63.7 [54.1–72.6]), proximal (81.5 [74.6–87.1] vs. 63.8 [55.9–71.2]), middle (81.2 [75.7–85.9] vs. 59.4 [53.0–65.6]) and distal stenosis location (67.4 [57.0–76.6] vs. 51.6 [41.1–62.0]).ConclusionIn a multicenter cohort with high disease prevalence, ML-CT-FFR offered improved diagnostic performance over cCTA for detecting hemodynamically relevant stenoses regardless of their location.     

基于交叉对比神经网络的心音分类

目的:通过交叉对比神经网络（CCNN）实现心音信号的自动分类，从而对心血管疾病进行早期诊断。方法:实验基于PhysioNet/Cinc 2016心音数据库。训练集和测试集数据来自互斥的健康受试者/病理患者，并以4:1的比例进行划分，输入CCNN。CCNN利用深度卷积神经网络进行特征提取，结合基于信息的相似度度量理论（IBS），对特征向量间的相似性进行度量并分类。结果:实验结果得出灵敏度为0.834 6，特异性为0.962 3，最终大赛综合得分为0.898 5。结论:CCNN使用交叉对比的输入模式扩充数据量，引入信号间的对比信息，同时在神经网络的训练过程中应用统计学思想，使网络具备良好的泛化性，更加适应医学数据量较少的场景，在心音分类中取得较好的结果。     

基于MDT的对分课堂联合CBL教学模式在软组织肿瘤临床教学中的应用

目的:探讨基于MDT的对分课堂联合CBL教学模式在软组织肿瘤临床教学中的应用效果及学员对该教学模式的评价。方法:选取2017年1月至2019年10月在我科接受住院医师规范化培训的50名学员为研究对象。随机分为对照组和实验组，每组25人。对照组授课方式为传统教学法，实验组授课采用MDT联合对分课堂和CBL教学法。教学结束后，采用命题考试进行教学效果考核；采用问卷调查的方法评估学生对教学模式的满意度。结果:实验组的选择题、简答论述、病例分析和总成绩分别为26.36±2.75、18.24±2.40、30.76±3.09、75.36±5.96，而对照组分别为24.40±3.80、16.60±2.10、29.04±2.86、70.04±6.30，两组比较有统计学差异（P<0.05）。在提高课堂学习效率、学习兴趣、自学能力、理论知识的理解和记忆能力，扩充专业知识，提高文献检索能力、分析问题和解决问题的能力、临床思维能力，这8个维度的赞成度调查中，实验组均优于对照组（P<0.05）。实验组学员对本组教学模式的接受度更高（P<0.05），但也有更多的实验组学员认为本组教学模式增加了学习负担（P<0.05）。结论:MDT联合对分课堂和CBL的教学模式应用于软组织肿瘤临床教学中，有利于提高教学效果，且接受度更高。     

An evaluation of wayfinding abilities in adolescent and young adult males with autism spectrum disorder

BackgroundWayfinding refers to traveling from place to place in the environment. Despite some research headway, it remains unclear whether individuals with Autism Spectrum Disorder (ASD) show strengths, weaknesses, or similarities in wayfinding compared with ability-matched typically developing (TD) controls.MethodThe current study tested 24 individuals with ASD, 24 mental-ability (MA) matched TD (MA-TD) controls, and 24 chronological-age (CA) matched TD (CA-TD) controls. Participants completed a route learning task and a survey learning task, both programmed in virtual environments, and a perspective taking task. Their parents completed questionnaires assessing their children’s everyday wayfinding activities and competence.ResultsOverall, CA-TD controls performed better than both the ASD group and the MA-TD group in both wayfinding tasks and the perspective taking task. Individuals with ASD performed similarly to the MA- TD controls on wayfinding performance except for backtracking routes. Perspective taking presented an area of deficit for people with ASD and it predicted individual differences in route learning and survey learning. Parents’ reports did not predict their children’s wayfinding performance. Two mini meta-analyses, including previous studies and the current study, showed a significant deficit in route learning, but not in survey learning for the ASD group relative to MA-TD controls.ConclusionsAlthough participants with ASD showed impairments in wayfinding relative to CA-TD controls, the impairment is not specific to their ASD, but rather due to their mental age. Nevertheless, route reversal in route learning may present unique difficulty for people with ASD beyond the effects of mental age.     

AI-assisted superresolution cosmological simulations

Cosmological simulations of galaxy formation are limited by finite computational resources. We draw from the ongoing rapid advances in artificial intelligence (AI; specifically deep learning) to address this problem. Neural networks have been developed to learn from high-resolution (HR) image data and then make accurate superresolution (SR) versions of different low-resolution (LR) images. We apply such techniques to LR cosmological N-body simulations, generating SR versions. Specifically, we are able to enhance the simulation resolution by generating 512 times more particles and predicting their displacements from the initial positions. Therefore, our results can be viewed as simulation realizations themselves, rather than projections, e.g., to their density fields. Furthermore, the generation process is stochastic, enabling us to sample the small-scale modes conditioning on the large-scale environment. Our model learns from only 16 pairs of small-volume LR-HR simulations and is then able to generate SR simulations that successfully reproduce the HR matter power spectrum to percent level up to 16 h−1Mpc and the HR halo mass function to within 10% down to 1011 M⊙. We successfully deploy the model in a box 1,000 times larger than the training simulation box, showing that high-resolution mock surveys can be generated rapidly. We conclude that AI assistance has the potential to revolutionize modeling of small-scale galaxy-formation physics in large cosmological volumes.

As telescopes and satellites become more powerful, observational data on galaxies, quasars, and the matter in intergalactic space becomes more detailed and covers a greater range of epochs and environments in the Universe. Our cosmological simulations (see, e.g., ref. 1) must also become more detailed and more wide-ranging in order to make predictions and test the effects of different physical processes and different dark-matter candidates. Even with supercomputers, we are forced to decide whether to maximize either resolution or volume, or else compromise on both. These limitations can be overcome through the development of methods that leverage techniques from the artificial intelligence (AI) revolution (see, e.g., ref. 2) and make superresolution (SR) simulations possible. In the present work, we begin to explore this possibility, combining knowledge and existing superscalable codes for petascale-plus cosmological simulations (3) with machine learning (ML) techniques to effectively create representative volumes of the Universe that incorporate information from higher-resolution models of galaxy formation. Our first attempts, presented here, involve simulations with dark matter and gravity only, and extensions to full hydrodynamics will follow. This hybrid approach, which will imply offloading simulations to neural networks (NNs) and other ML algorithms, has the promise to enable the prediction of quasar, supermassive black hole, and galaxy properties in a way that is statistically identical to full hydrodynamic models, but with a significant speed-up.Adding details to images below the resolution scale (SR image enhancement) has become possible with the latest advances in deep learning (DL; ML with NN; ref. 4), including generative adversarial networks (GANs; ref. 5). The technique has applications in many fields, from microscopy to law enforcement (6). It has been used for observational astronomical images by (7), to recover galaxy features from below the resolution scale in degraded Hubble Space Telescope images. Besides SR image enhancement, DL has started to find applications in cosmological simulations. For example, refs. 8 and 9 showed how NNs can predict the nonlinear formation of structures given simple linear theory predictions. NN models have also been trained to predict galaxies (10, 11) and 21-cm emission from neutral hydrogen (12) from simulations that only contain dark matter. GANs have been used in ref. 13 to generate image slices of cosmological models and to generate dark-matter halos from density fields (14). ML techniques other than DL find many applications, too. For example, Kamdar et al. (15) have applied extremely randomized trees to dark-matter simulations to predict hydrodynamic galaxy properties.Generating mocks for future sky surveys requires large volumes and high accuracy, a task that quickly becomes computationally prohibitive. To alleviate the cost, recently, Dai and Seljak (16) developed a Lagrangian-based parametric ML model to predict various hydrodynamical outputs from the dark-matter density field. In other work, Dai et al. (17, 18) sharpened the particle distribution using a potential gradient descent method starting from low-resolution (LR) simulations. Note, however, that these approaches did not aim to enhance the spatial or mass resolution of a simulation.On the DL side, recently, Ramanah et al. (19) explored using the SR technique to map density fields of LR cosmological simulations to that of the high-resolution (HR) ones. While the goal is similar, our work has the following three key differences. First, instead of focusing on the dark-matter density field, we aim to enhance the number of particles and predict their displacements, from which the density fields can be inferred. This approach allows us to preserve the particle nature of the N-body simulations and therefore to interpret the SR outputs as simulations themselves. Second, we test our technique at a higher SR ratio. Compared to ref. 19, which increased the number of Eulerian voxels by 8 times, we increase the number of particles and thus the mass resolution by a factor of 512. Finally, to facilitate future applications of SR on hydrodynamic simulations in representative volumes, we test our method at much smaller scales and in large simulations whose volume is much bigger than that of the training data.     

Identifying juvenile myoclonic epilepsy via diffusion tensor imaging using machine learning analysis

The aim of this study was to evaluate the feasibility of using a machine learning approach based on diffusion tensor imaging (DTI) to identify patients with juvenile myoclonic epilepsy. We analyzed the usefulness of combining conventional DTI measures and structural connectomic profiles. This retrospective study was conducted at a tertiary hospital. We enrolled 55 patients with juvenile myoclonic epilepsy. All of the subjects underwent DTI from January 2017 to March 2020. We also enrolled 58 healthy subjects as a normal control group. We extracted conventional DTI measures and structural connectomic DTI profiles. We employed the support vector machines (SVM) algorithm to classify patients with juvenile myoclonic epilepsy and healthy subjects based on the conventional DTI measures and structural connectomic profiles. The SVM classifier based on conventional DTI measures had an accuracy of 68.1% and an area under the curve (AUC) of 0.682. Another SVM classifier based on the structural connectomic profiles demonstrated an accuracy of 72.7% and an AUC of 0.727. The SVM classifier based on combining the conventional DTI measures and structural connectomic profiles had an accuracy of 81.8% and an AUC of 0.818. DTI using machine learning is useful for classifying patients with juvenile myoclonic epilepsy and healthy subjects. Combining both the conventional DTI measures and structural connectomic profiles results in a better classification performance than using conventional DTI measures or the structural connectomic profiles alone to identify juvenile myoclonic epilepsy.     

Rethinking PICO in the Machine Learning Era: ML-PICO

Background  Machine learning (ML) has captured the attention of many clinicians who may not have formal training in this area but are otherwise increasingly exposed to ML literature that may be relevant to their clinical specialties. ML papers that follow an outcomes-based research format can be assessed using clinical research appraisal frameworks such as PICO (Population, Intervention, Comparison, Outcome). However, the PICO frameworks strain when applied to ML papers that create new ML models, which are akin to diagnostic tests. There is a need for a new framework to help assess such papers. Objective  We propose a new framework to help clinicians systematically read and evaluate medical ML papers whose aim is to create a new ML model: ML-PICO (Machine Learning, Population, Identification, Crosscheck, Outcomes). We describe how the ML-PICO framework can be applied toward appraising literature describing ML models for health care. Conclusion  The relevance of ML to practitioners of clinical medicine is steadily increasing with a growing body of literature. Therefore, it is increasingly important for clinicians to be familiar with how to assess and best utilize these tools. In this paper we have described a practical framework on how to read ML papers that create a new ML model (or diagnostic test): ML-PICO. We hope that this can be used by clinicians to better evaluate the quality and utility of ML papers.     

影像后处理技术在神经外科教学查房中的应用

目的  评价以问题为基础的教学法(problem-based learning,PBL)联合影像后处理技术在神经外科教学查房中的应用效果。方法  选择神经外科典型病例，以病例为基础，以问题为导向布置课前预习内容，联合影像后处理技术在教学查房中以导向问题为主线实施教学培训，与传统教学方法相比评价教学效果。结果  PBL联合影像后处理技术教学查房模式和传统讲授教学查房模式的学员理论考试成绩分别为92.53±2.17分、81.40±6.17分，临床诊疗考核成绩分别为91.10±1.97分、79.60±3.27分，P值分别为<0.001、<0.001，差异有统计学意义。接受PBL联合影像后处理技术教学模式的学员对教学方式的认可度、增加学习兴趣、利于神经科思维，P值分别为0.013、0.024、0.035。结论  PBL联合影像后处理技术可有效提高教学查房效果，帮助学员提高对基础知识的理解能力，抓住神经外科疾病的关键点加速神经外科诊疗思维能力的形成。