Deep Learning in Healthcare
With successful experimental results and wide applications, Deep Learning (DL) has the potential to change the future of healthcare. The use of Artificial Intelligence (AI) has become increasingly popular and is now used, for example, in cancer diagnosis and treatment. It is thus no surprise that a recent report from ReportLinker has noted that the AI healthcare market is expected to grow from $2.1 billion in 2018 to $36 billion by 2025. Let’s see more about the potential of deep learning in the healthcare industry and its many applications in this field. In this article:
- Deep learning: the future of healthcare
- The use of deep learning in electronic health records
- Deep learning applications in healthcare
- Deep learning for cancer diagnosis
- Deep learning in disease prediction and treatment
- Scaling up deep learning in healthcare with MissingLink
We have used Artificial Intelligence (AI), in the traditional sense, and algorithmic learning to help us understand medical data, including images, since the initial days of computing. Today’s interest in Deep Learning (DL) in healthcare is driven by two factors. First, the growth of deep learning techniques, in the broad sense, and particularly unsupervised learning techniques, in the commercial area with, for example, Facebook, Google, and IBM Watson. Second, the dramatic increase of healthcare data that stems from the HITECH portion of the American Recovery and Reinvestment Act (ARRA). The latter worked to change records from carbon paper to silicon chips, in the form of unstructured, structured and available data.
Electronic Health Record (EHR) systems store patient data, such as demographic information, medical history records, and lab results. EHR systems improve the rate of correct diagnosis and the time it takes to reach a prognosis, via the use of deep learning algorithms. These algorithms use data stored in EHR systems to detect patterns in health trends and risk factors and draw conclusions based on the patterns they identify.
Researchers can use data in EHR systems to create deep learning models that will predict the likelihood of certain health-related outcomes such as the probability that a patient will contract a disease.
Two ways to use EHR system data:
1. A static prediction A static prediction, tells us the likelihood of an event based on a data set researchers feed into the system and code embeddings from the International Statistical Classification of Diseases and Related Health Problems (ICD). For example, Choi et al. fed a DL model with the representation of a patient created from EHR data, specifically, their medical history and their rate of hospital visits. Based on this information, the system predicted the probability that the patient will experience heart failure. 2. A prediction based on a set of inputs Data from the EHR system is used to make a prediction based on a set of inputs. It is possible to either make a prediction with each input or with the entire data set. For example, Choi et al. developed Doctor AI, a model that uses Artificial Neural Networks (ANN) to predict when a future hospital visit will take place, and the reason prompting the visit. They base this prediction on the information including, ICD codes gathered from a patient’s previous hospital visits and the time elapsed since the patient’s most recent visit.
Generative Adversarial Network (GAN)
Using EHR data is difficult in a scenario when doctors are required to diagnose rare diseases or perform unique medical procedures with little available data. To solve this issue, doctors and researchers use a deep learning method called Generative Adversarial Network (GAN). In the following example, the GAN uses data from patients records and creates more datasets, which the model trains on. GAN pits two rivaling ANNs against each other, one is called a generator and the other a discriminator, within the same framework of a zero-sum game. The generator will learn the specifics of a given dataset and will generate new data instances in an attempt to fool the discriminator into thinking they are genuine. Then, the discriminator will test both data sets for authenticity and decide which are real (1) and which are fake (0). This process repeats, forcing the generator to keep training in an attempt to produce better quality data for the model to work with.
Deep learning techniques use data stored in EHR records to address many needed healthcare concerns like reducing the rate of misdiagnosis and predicting the outcome of procedures. By processing large amounts of data from various sources like medical imaging, ANNs can help physicians analyze information and detect multiple conditions:
- Analyze blood samples
- Track glucose levels in diabetic patients
- Detect heart problems
- Using image analysis to detect tumors
- Detecting cancerous cells and diagnosing cancer
- Detecting osteoarthritis from an MRI scan before the damage has begun
Oncologists have been using methods of medical imaging like Computed Tomography (CT), Magnetic Resonance Imaging (MRI) and X-ray to diagnose cancer for many years. While these systems have proven to be effective for many types of cancer, a large number of patients suffer from forms of cancer that cannot be accurately diagnosed with these machines. ANNs like Convolutional Neural Networks (CNN), a class of deep learning, are showing promise in relation to the future of cancer detection. Based on the same medical images ANNs are able to detect cancer at earlier stages with less misdiagnosis, providing better outcomes for patients. Recently, scientists succeeded in training various deep learning models to detect different kinds of cancer with high accuracy.
- In a study published by Nvidia, a deep learning model was able to decrease the misdiagnosis rate of breast cancer by 85%.
- Inspired by his roommate, who was diagnosed with leukemia, Hossam Haick attempted to create a device that treats cancer. Based on his design, a team of scientists trained an ANN model to identify 17 different diseases based on patients smell of breath with 86% accuracy.
- A team of researchers at Enlitic introduced a device that surpassed the combined abilities of a group of expert radiologists at detecting lung cancer nodules in CT images, achieving a 50% higher detection rate in testing conditions.
- Haenslle et al. trained a CNN model to diagnose skin cancer by determining if a skin lesion shown in digital imaging is cancerous with the accuracy of an expert dermatologist.
- Scientists at Google have created a CNN model that detects metastasized breast cancer from pathology images faster and with improved accuracy. LYmph Node Assistant (LYNA), achieved a 99% success rate, compared to human doctors who scored as low as 38% in relation to some individual test slides.
- Aidoc’s AI for radiology images analysis is installed in more than 100 hospitals worldwide and helped radiologists treat over a million patients. After an initial struggle with resource management and inflated GPU cloud service costs, Aidoc started using MissingLink.ia with success. With MissingLink’s automated platform, Aidoc’s AI team could scale up as they planned. The platform lets Aidoc’s team automate and control their deep learning lifecycle, their core cloud infrastructure, and their experiment results. Furthermore, Aidoc’s AI team can use MissingLink to view and control their experiments, aggregated from multiple cloud servers, from a single web page. The team can also queue and prioritize experiment while MissingLink automatically manages deployment.
In 2006, over 4.4 million preventable hospitalizations cost the U.S. more than $30 billion. Half of the patients hospitalized suffer from two conditions: heart problems and diabetes. Deep learning can be used to improve the diagnosis rate and the time it takes to form a prognosis, which may drastically reduce these hospitalization numbers. Some research teams are already applying their solutions to this problem:
- A team of Researchers from Boston University collaborated with local Boston hospitals. They monitor and predict with 82% accuracy which patients with heart disease and diabetes will require hospital care in the next year.
- Researchers created a medical concept that uses deep learning to analyze data stored in EHR and predict heart failures up to nine months before doctors can.
Diabetic Retinopathy (DR)
In developing countries, more than 415 million people suffer from a form of blindness called Diabetic Retinopathy (DR), which is caused by complications resulting from diabetes. Deep learning can help prevent this condition. A CNN model can work with data taken from retinal imaging and detect hemorrhages, the early symptoms, and indicators of DR. Diabetic patients suffer from DR due to extreme changes in blood glucose levels. A team of scientists suggests that diabetic patients can be monitored for their glucose levels. A deep learning model can use this data to predict when these spikes or drops will occur, allowing patients to respond by either eating a high-sugar snack or injecting insulin.
Human Immunodeficiency Virus (HIV)
Over 36 million people worldwide suffer from Human Immunodeficiency Virus (HIV). These individuals require daily doses of antiretroviral drugs to treat their condition. HIV can rapidly mutate. Thus to keep treating HIV, we must keep changing the drugs we administer to patients. Using a Deep learning model called Reinforcement Learning (RL) can help us stay ahead of the virus. In this HIV scenario, the RL model (the agent) can track many biomarkers (the environment) with every drug administration and provide the best course of action to alter the drug sequence for continuous treatment. DeepBind: Genome Research Understanding our genomes can help researchers discover the underlying mechanisms of diseases and develop cures. A team of researchers at the University of Toronto have created a tool called DeepBind, a CNN model which takes genomic data and predicts the sequence of DNA and RNA binding proteins. Researchers can use DeepBind to create computer models that will reveal the effects of changes in the DNA sequence. They can apply this information to develop more advanced diagnostic tools and medications.
Privacy Issues arising from using Deep Learning in Healthcare
Despite the many advantages of using large amounts of data stored in patients EHR systems, there are still risks involved. The data EHR systems store also contains personal information many people prefer to keep private like previous drug usage. Hospitals also store non-medical data such as patients addresses and credit card information which makes these systems a primary target for attacks from bad actors. With the amount of sensitive data stored in EHR and its vulnerability, it is critical to protect it and keep the patients’ privacy. This can be done with MissingLink data management.
Using deep learning in healthcare typically involves intensive tasks like training ANN models to analyze large amounts of data from many images or videos. Running these models demand powerful hardware, which can prove challenging, especially at production scales. Using MissingLink can help by providing a platform to easily manage multiple experiments. Schedule, automate and record your experiments and save time and money. Learn more and see how easy it is to use deep learning in healthcare with MissingLink.