周读论文系列笔记(1)-review-Artificial intelligence in healthcare

版权声明：本文为博主原创文章，未经博主允许不得转载。 https://blog.csdn.net/Void_worker/article/details/85290027

第一次…不知道怎么写，好纠结(꒦_꒦)

原文链接：https://www.nature.com/articles/s41551-018-0305-z

1.A historical overview of AI in medicine

(1)First generation of AI systems : clinical decision support systems(mid-twentieth century:)
Rule-based approaches. to interpret ECGs, diagnose diseases, choose appropriate treatments, provide interpretations of clinical reasoning and assist physicians in generating diagnostic hypotheses in complex patient cases.
(2)machine learning
(3)deep learning

2.Recent breakthroughs in AI technologies and their biomedical applications

2.1 Image-based diagnosis 基于图像的诊断

Currently, automated medical-image diagnosis is arguably the most successful domain of medical AI applications.

2.1.1 Radiology 放射学

To use medical-imaging modalities to detect and diagnose diseases.
medical-imaging modalities: X-ray radiography, computed tomography, magnetic resonance imaging(MRI) and positron-emission tomography
Radiological practice relies primarily on imaging for diagnosis.

With the help of modern machine-learning methods, many radiology applications of AI, such as the detection of lung nodules using computed tomography images(使用计算机断层扫描肺结节), the diagnosis of pulmonary tuberculosis(肺结核) and common lung diseases with chest radiography(胸部X线检查的常见肺部疾病) and breast-mass identification using mammography scans(乳房X射线摄影扫描的乳房肿块), have reached expert-level diagnostic accuracies.

These studies employed a technique known as transfer learning, where well-established deep neural networks trained on millions of natural, non-medical images are borrowed, and then the neural network connections are finetuned by using thousands of biomedical images.[使用迁移学习]

Regulatory approval: a deep-learning system for diagnosing cardiovascular diseases(心血管疾病) using cardiac MRI images was approved by the FDA in 2018.[监管部门的批准]

2.1.2 Dermatology 皮肤病学

typical skin melanoma（典型的皮肤黑色毒瘤）
pigmented tumours（色素性肿瘤）ABCED rule.
classify photographs of benign and malignant lesions（对良性和恶性病变的分类） CNN trained on 129,450 clinical images achieved dermatologist-level accuracy in diagnosing skin malignancy.

2.1.3 Ophthalmology 眼科学

Fundus photography（眼底照相）
It can detect and monitor diseases such as DR, glaucoma（青光眼）, neoplasms of the retina（视网膜肿瘤） and age-related macular degeneration（年龄相关性黄斑疾病）, and plays a vital role in identifying causes of preventable blindness（可预防性失明）
CNN trained to identify referable DR and diabetic macular oedema（糖尿病黄斑水肿） using 128,175 retinal images.
Deep learning can extract unrecognized associations between retinal image patterns and age, gender, systolic blood pressure and smoking status, as well as major adverse cardiac events.
Another team of researchers showed that the performance of a CNN exceeded their pre-specified sensitivity (85%) and specificity (82.5%); the system was authorized by the FDA for use by healthcare providers to detect diabetic macular oedema（糖尿病黄斑水肿） and moderate-to-severe DR（中度至重度DR）

2.1.4 Pathology 病理学

Histopathological（组织病理学） assessment is the gold standard for the diagnosis of many cancer types.
With CNN, AI can be useful in the detection of prostate cancer（前列腺癌） from biopsy specimens（活组织切片）, the identification of breast cancer metastasis in lymph nodes（乳腺癌淋巴结转移） and the detection of mitosis in breast cancer（乳腺癌有丝分裂）.

2.2 Genome interpretation 基因组解释

High-throughput sequencing methods generate terabytes of raw data for genomic studies.
DNN can annotate pathogenic genetic variants（致病性遗传变异） and identify the functions of non-coding DNA（识别非编码DNA的功能）.

2.3 Machine learning for biomarker discovery 生物标志物的发现

Biomarker discovery relies on identifying previously unrecognized correlations between thousands of measurements and phenotypes.
Machine learning can identify the molecular patterns associated with disease status and disease subtypes, account for the high-level interactions among the measurements, and derive omics signatures to predict disease phenotypes.
cancers, infectious diseases and the risk of Down’s syndrome（唐氏综合征）.

2.4 Clinical outcome prediction and patient monitoring 临床结果预测和患者监测

The use of EHRs to predict clinical outcomes. 使用EHR预测临床结果
Bayesian networks can predict mortality（死亡）, readmission（再住院率） and length of hospital stay（住院时间）.
Classify cancer patients with different responses to chemotherapy（对具有不同化疗反映的癌症患者分类）, and clinical predictors for the prognosis of patients receiving thoracic organ transplantation（接受胸腔器官移植） can be identified.
Routine monitoring devices generate a large amount of data and thus represent a great opportunity for AI-assisted alert systems.
a prediction model for cardiac arrest（心脏骤停预测模型）
assist anesthesiologists in predicting hypoxemia events during surgery（预测手术过程中的低氧血事件）.

2.5 Inferring health status through wearable devices 通过可穿戴设备推测健康状态

2.6 Autonomous robotic surgery 自主机器人手术

3.Technical challenges in AI developments

4.Social, economic and legal challenges