机器学习与分布式机器学习
Machine learning (ML) has spread into many different fields and disciplines. Dipping your toes into a new field is the best way to grow and learn new things. The following is a summary of how researchers have applied machine learning to diagnose and treat Autism Spectrum Disorder (ASD).
机器学习(ML)已扩展到许多不同的领域和学科。 将脚趾浸入新领域是成长和学习新事物的最佳方法。 以下是研究人员如何将机器学习应用于诊断和治疗自闭症谱系障碍(ASD)的摘要。
论文(按顺序) (Papers (In order))
All of these papers are accessible without any university sponsorship or payment.
所有这些论文都可以在没有大学赞助或任何付款的情况下获得。
Personalized machine learning for robot perception of affect and engagement in autism therapy
Use of machine learning to shorten observation-based screening and diagnosis of autism
眼动与自闭症 (Eye Movement and Autism)
The authors, Wenbo Liu, Ming Li, and Li Yi, developed an ML model to identify ASD based on eye movements when viewing faces.
作者，刘文博，李明和李毅，开发了一种ML模型，该模型可根据观看面部时的眼动识别ASD。
Eye-tracking, SVM, RBF Kernal, References paper 2 DP Wall, J. Kosmicki
眼动追踪，SVM，RBF Kernal，参考文献2 DP Wall，J。Kosmicki
思想 (Ideology)
Individuals with ASD have difficulty recognizing human faces and interpreting facial emotions. Additionally, individuals with ASD often have an atypical face-scanning pattern. The authors wanted to know if you can identify ASD based on where and how long a person looks when viewing a face.
患有自闭症的人很难识别人脸和解释面部表情。 此外，患有自闭症的人通常具有非典型的面部扫描模式。 作者想知道您是否可以根据一个人在看脸时看到的位置和时间长短来识别ASD。
数据 (Data)
The individuals in the study are 29 children with ASD and 29 children without ASD. The data used in this paper was collected from a different study where researchers looked at if children with ASD had a hard time identifying people of different races.
研究对象为29名患有ASD的儿童和29名没有ASD的儿童。 本文使用的数据来自另一项研究，研究人员研究了ASD儿童是否很难识别不同种族的人。
方法 (Methods)
Researchers divided up every face into K different sections using a K-means algorithm. They then produced histograms measuring how often individuals looked at each distinct cell. This information was used to train an SVM to predict if the viewer has ASD or not.
研究人员使用K-means算法将每张脸分成K个不同的部分。 然后，他们产生了直方图，用于测量个人查看每个不同细胞的频率。 此信息用于训练SVM以预测查看器是否具有ASD。
结果 (Results)
The SVM was able to predict if a viewer has ASD or not with 88.51% accuracy. Although not impervious to false positives and false negatives, the model does support that lack of eye contact is a strong sign of ASD.
SVM能够以88.51％的准确性预测观众是否具有ASD。 该模型虽然不能完全避免误报和误报，但确实支持缺乏眼神交流是ASD的强烈信号。
机器人增强疗法 (Robot-enhanced Therapy)
The MIT Media Lab put a robot inside the clinic to assist with autism therapy. Additionally, they made specialized models for each kid that interacted with it.
麻省理工学院媒体实验室在诊所内放置了一个机器人来协助自闭症治疗。 此外，他们为与之互动的每个孩子制作了专门的模型。
Robotics, Therapy, T-SNE, DeepLift, OpenPose, OpenFace, openSMILE
机器人技术，治疗，T-SNE，DeepLift，OpenPose，OpenFace，openSMILE
思想 (Ideology)
Robots have been used for autism therapy in the past but its effectiveness could be debated. Each child with autism is different.
过去，机器人已用于自闭症治疗，但其有效性尚有争议。 每个患有自闭症的孩子都是不同的。
“If you have met one child with autism, you have met one child with autism.”
“如果您遇到一个患有自闭症的孩子，您就遇到了一个患有自闭症的孩子。”
Thus, generic models for predicting emotions won’t work. The solution is to make a specialized model for each child that interacts with the robot. That way the robots know what features constitute happy or sad.
因此，用于预测情绪的通用模型将不起作用。 解决方案是为每个与机器人互动的孩子建立专门的模型。 这样，机器人就知道哪些特征构成快乐或悲伤。
方法 (Methods)
The researchers had each child play in a room with a clinician while wearing a watch to keep track of physiological data. The clinician would record the child’s emotions while playing. This data, along with video and audio, was used to train the model for the robot.
研究人员让每个孩子都在与医生一起在房间里玩耍，同时戴着手表来跟踪生理数据。 临床医生会在玩耍时记录孩子的情绪。 这些数据以及视频和音频被用来训练机器人的模型。
影响 (Impacts)
Previously, when a robot was deployed in a clinical setting, it mimicked emotions that the therapist described or told the child how it felt during a story. Now, the robot can respond to a child’s emotions and prompt suggestions/questions like, “Why are you sad?” or “Let’s take a break.”
以前，在临床环境中部署机器人时，它模仿的是治疗师描述或告诉孩子在故事中的感受的情绪。 现在，机器人可以响应孩子的情绪并提示诸如“你为什么难过？”之类的建议/问题。 或“让我们休息一下。”
计算机视觉辅助ASD诊断 (Computer vision aided ASD diagnosis)
Researchers from Duke University designed a low cost and non-invasive Computer Vision (CV) tool to help diagnose autism.
杜克大学的研究人员设计了一种低成本且无创的计算机视觉(CV)工具来帮助诊断自闭症。
AOSI, Interrater score, Tracking facial features
AOSI，Interrater得分，跟踪面部特征
方法 (Methods)
The study focused on at-risk children (children who have a sibling diagnosed with ASD) who were between 7–14 months. All of the children were evaluated by a professional clinician using the Autism Observational Scale for Infants (AOSI). Video from this evaluation was seen by one specialized AOSI clinician, one general psychologist, and two psychology students. They all rated the children based on how the child reacted to different auditory stimuli.
这项研究的重点是7至14个月之间的高危儿童(被诊断患有ASD的兄弟姐妹的儿童)。 由专业临床医生使用婴儿自闭症婴儿观察量表(AOSI)对所有儿童进行评估。 一位专业的AOSI临床医生，一位普通的心理学家和两名心理学专业的学生看了这次评估的视频。 他们都根据孩子对不同听觉刺激的React对孩子进行评估。
The researchers then automated the review process by using CV tools to track facial movements as the child looked at the different stimuli. After measuring the child’s responses, they would assign the same ‘pass’, ‘delayed’, and ‘stuck’ scores as the human evaluators.
然后，研究人员通过使用CV工具跟踪孩子观察不同刺激时的面部运动，从而使检查过程自动化。 在测量了孩子的回答后，他们将分配与人类评估者相同的“及格”，“延迟”和“卡住”分数。
结果 (Results)
By only tracking eyes, noes, and ears the automated method agreed with the professional AOSI clinician 89% of the time and received an excellent interrater score of 75%.
仅通过跟踪眼睛，鼻子和耳朵，自动化方法就可以在89％的时间内与专业的AOSI临床医生达成一致，并获得了75％的出色间质评分。
影响 (Impacts)
This research should be used as a stepping stone for more CV aided ASD diagnosis tools. The original work is very limiting but can be expanded to allow for different clinical tests or even home videos. The ultimate goal is to bring down the cost of ASD diagnosis and this work contributes to that.
这项研究应作为更多的CV辅助ASD诊断工具的垫脚石。 原始作品非常有局限性，但可以扩展以允许进行不同的临床测试甚至家庭录像。 最终目标是降低ASD诊断的成本，这项工作为此做出了贡献。
ADTree缩短了基于观察的ASD筛选 (ADTree Shortens Observation Based ASD Screening)
Researchers shortened the most widely used instrument for evaluation of ASD, Autism Diagnostic Observation Schedule-Generic (ADOS), from 29 items to 9 items without appreciable loss of accuracy.
研究人员将最广泛使用的ASD评估工具，即自闭症诊断观察计划表(ADOS)，从29项缩短为9项，而没有明显的准确性损失。
ADOS, ADTree, Upsampling
ADOS，ADTree，上采样
数据 (The Data)
The researchers obtained 1073 total ADOS observations from Boston Autism Consortium (AC), Autism Genetic Resource Exchange(AGRE), and the Simons Foundation. Out of the 1073 observations, only 15 of them come from individuals without autism. Some of the observations were missing data and thus not used. To account for the imbalanced classes researches upsampled from 15 observations to 612.
研究人员从波士顿自闭症协会(AC)，自闭症遗传资源交易所(AGRE)和西蒙斯基金会(Simons Foundation)获得了总计1073份ADOS观测值。 在1073项观察中，只有15项来自没有自闭症的个体。 一些观察缺少数据，因此未使用。 为了解决类别不平衡的问题，研究从15个观测值上采样到612个。
模型 (The Models)
The paper focuses on creating a machine learning model that can help clinicians quantify results from ADOS evaluations. During this process, they developed 16 different models that all had varying levels of success.
本文着重于创建一种可以帮助临床医生量化ADOS评估结果的机器学习模型。 在此过程中，他们开发了16种不同的模型，每个模型都有不同的成功水平。
ADTree model preformed the best with 100% accuracy. Along with correctly classifying every sample given the ADTree pruned away many of the inputs given. The ADTree only needs 9 of the original 27 items to achieve the same level of accuracy.
ADTree模型以100％的准确度预演了最好的模型。 在给定给定的每个样本的正确分类的同时，ADTree删除了给定的许多输入。 ADTree只需要原始27个项目中的9个就可以达到相同的精度水平。
影响 (Impacts)
This reduction reduces the amount of work by a clinician by 66% and therefore speeds up the ASD diagnoses pipeline. With more ADOS data reported and logged ADTrees may be able to categorize individuals into different areas on the spectrum.
这种减少使临床医生的工作量减少了66％，因此加快了ASD诊断流程。 通过报告和记录更多的ADOS数据，ADTree可以将个人分类到频谱的不同区域。
机器学习和自闭症的下一步是什么？ (What’s next for machine learning and autism?)
It is exciting to see all of the research being done in the early detection of autism as well as new therapies for children. I enjoyed the paper about ADTrees shortening observation-based ASD screening the most and think that strategies like tree pruning could lead to more effective surveys for other diseases as well.
令人兴奋的是，在自闭症的早期发现以及针对儿童的新疗法方面正在进行的所有研究都令人兴奋。 我最喜欢关于ADTrees缩短基于观察的ASD筛选的论文，并认为像树修剪这样的策略也可以导致对其他疾病的更有效调查。
I would love to see more work done on DNA and autism. Autism speaks has a great article describing some of the work that has been done in that field.
我希望看到有关DNA和自闭症的更多工作。 自闭症的演讲有一篇很棒的文章，描述了该领域已经完成的一些工作。
翻译自: https://towardsdatascience.com/what-is-machine-learning-doing-for-autism-6e86daa3b12a
机器学习与分布式机器学习
扫一扫
1059
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
