Projects

Summary

Figure 1 | Innovative AI-enhanced clinical process. The ECGs could be real-time and automatically uploaded to the AI platform for autodiagnosis. When AI indicates fatal diseases, a smartphone notification message triggered by AI-based automatic active alarm system would be sent to front-line physicians and the on-duty cardiologist. All the physicians could prepare for subsequent confirmation and management.

Reference:
¹npj Digital Medicine, 5(1):8.
²Clin Chim Acta, 536:126-134.
³JMIR Med Inform, 8(3):e15931.
⁴Unpublished data.
⁵EuroIntervention, 17(9):765-773.
⁶J Pers Med, 11(11):1149.
⁷Unpublished data.
⁸Can J Cardiol, 38(2):160-168.
⁹J Pers Med, 12(7):1150.
¹⁰Unpublished data.
¹¹Eur J Trauma Emerg Surg, 48(4):3317-3326.
¹²Int J Environ Res Public Health, 18(7), 3839.
¹³J Endocr Soc, 5(9), bvab120.
¹⁴We shared our works to the Discovery.

Summary

Deep learning is a data-driven algorithm, so it is necessary to understand the limitations of medical data and design suitable algorithms to build an accurate medical artificial intelligence system. We consider that main differences between medical field and other fields and solutions are as follows:
1. Unsupervised learning aided supervised learning to solve the rare data problem: Because the medical data involves manual operations and expensive instruments, the amount of labeled data is scarce compared with other industries. Moreover, the incidence and prevalence of more fatel diseases is lower compared with other common diseases. We need to face this important limitation rather than selecting sufficient data for research. Transfer learning is a common solution in deep learning training that can effectively reduce the number of samples required. Our main work is to use a large amount of unlabeled data for unsupervised algorithms, and apply them to supervised learning.¹
2. Combination of multi-level statistical models and deep learning to create personalized artificial intelligence: The response of each indivial to the same treatment is different, so it is impossible to use one artificial intelligence models for accurately appling in everyone. Therefore, it is very important to establish a personalized artificial intelligence model, but it is difficult to train each person's data separately in medical field. The multi-level data analysis in statistics shows how to use the distribution assumptions to construct personalized predictions. Even if only one record also can be used to construct a personalized model, our research attempts to apply this concept to deep learning training. The new age medical artificial intelligence can dimamicly revise the prediction results.²
3. To design special prediction functions and loss functions: When using artificial intelligence models in the medical field, interpretability and confidence issues often arise. For example, if the data quality is poor, it may be advisable for frontline operators to perform the check again immediately. This is particularly important in prospective trials.³
4. Special accuracy evaluation metric: Medical artificial intelligence research often ignores the problem of variable dependence in medical data, such as the relationship between electrocardiograms and age and gender, and the relationship between gender and age and bone density, leading to confusion about whether the relationship between electrocardiograms and bone density is based on a real causal relationship. Because of the powerful feature extraction ability of deep learning models, AI currently often learns false correlations to establish predictive logic, which will reduce the accuracy of external validation in the future, so it is necessary to design appropriate evaluation indicators for this problem.⁴

Figure 2 | DLM training strategy for ECG data. To apply the DLM pre-trained by ImageNet for improving the accuracy is common in medical image analysis. We revolutionarily developed an unsupervised training strategy enhancing the feature extraction ability of DLM in ECGs through patient ID learning. This will improve the model accuracy especially in rare disease detections.

Reference:
¹Unpublished data.
²Eur Heart J-DH, ztac072.
³Digital Health, 8:20552076221143249.
⁴Unpublished data.

Summary

Participants

Chin Lin
Adviser

Yi-Ying Wu
Hematologist

Dung-Jang Tsai
Research fellow

Yu-Sheng Luo
PhD student

Ying-Chu Chen
MS student

Xin-An Lin
Engineer

Yu-Cheng Chen
Cardiologist

Hao-Chun Liao
BS student

Summary

Machine learning has seen some dramatic developments recently, leading to a lot of interest from industry, academia and popular culture. These are driven by breakthroughs in artificial neural networks, often termed deep learning, a set of techniques and algorithms that enable computers to discover complicated patterns in large data sets. Feeding the breakthroughs is the increased access to data (“big data”), user-friendly software frameworks, and an explosion of the available compute power, enabling the use of neural networks that are deeper than ever before. These models nowadays form the state-of-the-art approach to a wide variety of problems in computer vision, language modeling and robotics. Deep learning rose to its prominent position in computer vision when neural networks started outperforming other methods on several high-profile image analysis benchmarks. Most famously on the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) in 2012 when a deep learning model (a convolutional neural network) halved the second best error rate on the image classification task. Enabling computers to recognize objects in natural images was until recently thought to be a very difficult task, but by now convolutional neural networks have surpassed even human performance on the ILSVRC, and reached a level where the ILSVRC classification task is essentially solved (i.e. with error rate close to the Bayes rate). Deep learning techniques have become the de facto standard for a wide variety of computer vision problems. Healthcare providers generate and capture enormous amounts of data containing extremely valuable signals and information, at a pace far surpassing what “traditional” methods of analysis can process. Machine learning therefore quickly enters the picture, as it is one of the best ways to integrate, analyze and make predictions based on large, heterogeneous data sets. Recently, we have developed an object detection model for bone marrow smear interpretation¹. The chest X-ray can also be used in cardiovascular diesaese diagnosis, such as heart failure detection², etc.

Reference:
¹JMIR Med Inform, 8(4):e15963.
²Can J Cardiol, 38(6):763-773.

---

Medical Natural Language Processing

Participants

Chin Lin
Adviser

Wen-Hui Fang
Clinical expert

Dung-Jang Tsai
Research fellow

Yu-Sheng Luo
PhD student

Xin-An Lin
Engineer

Summary

Automated extrating structured medical information methods are increasingly being researched because of the increasing volume and accessibility of electronic medical data. However, it is difficult to use because more than 80% information is hidden in free-text clinical narratives. The current methods for collecting this structured information usually involve manual identification, but manual identification of disease codes from free-text clinical narratives is laborious and costly. The traditional document classification algorithm is to use bag-of-words models and machine learning meclassifiers, such as SVM and Random Forest. However, it is not really accurate and cannot face emerging diseases. Our team has developed an automatic disease classification approach by word embedding combining convolutional neural network, and it increases 30% performance compared with traditional methods¹. Moreover, it can handle emerging disease issue by external vocabulary resources. We can obtain more data-driven clues by Artificial Intelligence to help promote the progress of medicine. The health care field will then truly enter the age of big data².This technology was also applied in the medical record scoring system³.

Reference:
¹J Med Internet Res, 19(11):e380.
²JMIR Med Inform, 7(3):e14499.
³Healthcare, 9(10):1298.