Practical uses of Deep Neural Networks in Healthcare
Ray Kurzweil proposed that the rate of change in a wide variety of evolutionary systems tends to increase exponentially. This is especially true in the field of information technology. These technologies in isolation make impressive strides in their own way as they go through this growth spurt, but something fascinating and magical happens when they converge — this leads to industry disruption. The topic of this article is to take a look at how advances in Big Data, Machine Learning, Cloud Computing, and medical science have completely disrupted the Healthcare industry.
There is a deluge of data in the healthcare industry in the last few years that we accumulate every data ranging from EHR data, medical imaging, wearable devices, patient monitoring data etc. This rich data along with the increase in computational power and cloud computing has enabled us to achieve major breakthrough in certain areas of Healthcare. We will look at some of these areas of improvement and practical examples.
Supervised Learning: Addressing Delayed Intensive Care Unit (ICU) Admission
Identifying the right diagnosis for a given person at the earliest is one of the fundamental problems in healthcare faced by healthcare providers, unless of course, you are Dr. House M.D. Unfortunately, not every doctor can be Dr. House and secondly (and more importantly) Dr. House is fictional!
Timely diagnosis continues to elude the best of Doctors and every minute is critical for a patient during the triage process into Emergency care which can need a possible ICU admission. Delayed ICU admission is correlated with mortality. The underlying reason is ignoring correlations among vital signs, history and patient heterogeneity. Research has shown that Machine Learning is literally saving lives by scouring a multitude of patients’ data and comparing them to one patient’s health data to detect symptoms 12 to 24 hours sooner than a doctor.
I will share a real-life example to show how delayed ICU admission can turn out fatal in some cases. The patient in this example visits an emergency room to be treated for foot sore. The doctors inspect her, don’t see major concerns, and admit her in the general ward. Over the course of a few days, the patient develops Pneumonia, tachycardia (high heart rate), has trouble breathing, and her body enters into Septic shock. Incidentally, mortality in shock is 1 in 2. The patient gets transferred to ICU and the doctors try to stabilize her. The condition worsens and her kidneys and lungs stop functioning eventually leading to her death on Day 22.
Although the patient received the right treatments in this case, it was late unfortunately. Research has shown that for every hour treatment is delayed, mortality rate is increased 7–8%. Sepsis is the 11th leading cause of death more than the combined rate of Breast Cancer and Prostrate Cancer combined! It is preventable if treated early. It has been observed that Doctors have a really hard time diagnosing Sepsis. Coincidentally, this is where Machine Learning can help ingest all the EHR data to perform diagnosis. Since medical data is so huge and diverse, it is hard to correlate the different symptoms and remove the false positives. Because of machine learning algorithms being computationally stronger, they are able to understand this complex correlation earlier than humans as shown in the pictures below.
Moreover, algorithms can run 24/7 as compared to Doctors and it’s easier to scale up their computational power.
Convolution Neural Networks and Medical Imaging
Imaging is the fasting growing sources of medical data in the form of X-Ray, CAT Scan, MRI, Mammogram etc. The medical community realized that imaging can sometimes reveal a disease even before the symptoms are shown. For this reason, there are screenings as a preventive check such as Pap smears, Mammograms, Colonoscopy etc. For certain diseases like Cancer, early diagnosis increases the changes of survival and healthy life style and also reduces treatment costs. The challenge here is that a trained human eye of a medical professional cannot identify subtle signs on these tests. Machine learning (specifically Convolutional Neural Networks, CNN) has proven to be a powerful aid in this context. The algorithm is well suited for a multi-class classification problem (predicting picture as a car or a bicycle or a van) or binary classification (for example — predicting whether a medical image has a malignant tumor or not).
We will look at two examples where Image Recognition techniques can help detecting the diseases.
Osteoarthritis
Osteoarthritis is one of those diseases that cannot be detected until the damage is done and 1 in 10 people will develop it. Until now, Doctors couldn’t detect it until the bone damage has begun. Let’s take a look at the MRI scan of knee cartilage of two sets of patients.
The image on the left will be diagnosed with Osteoarthritis and the one on right will not over the next 3 years. The differences are subtle yet present, even for expert Doctors. With the help of Machine Learning, we are now able to input a training set of images of past patients who were diagnosed with the disease and healthy ones, the algorithm detects these patterns and is able to predict who is susceptible to disease in 3 years! This information was present in millions of scans taken till date but the human eye was not trained to detect it. Thankfully, the advances in Deep Learning and research in CNN has enabled to detect these subtleties in Images
Retinal Imaging
Diabetic Retinopathy (DR) is the fastest growing cause of blindness. There are 415 Million people with diabetes in the world today, and each one is potentially at risk of being diagnosed with DR. Key to prevent blindness is regular screening as DR does not show any symptoms till it’s too late to a point where there is vision loss. During the screening process, Doctors look for spots or hemorrhages to detect the presence of DR and grade them on a scale to show the health of patient’s eye.
In certain countries where there are not enough Doctors, a big portion of the diabetic population suffer from blindness even before they are diagnosed. This is totally preventable. There is another challenge even in a developed country like the United States.
The example above shows a patient data who are screened (one row per patient). The columns indicate diagnosis rating by each Doctor. This data was collected using US Board certified Ophthalmologists. As you can see at the extremes (no disease or end stage at the top and bottom respectively), the doctors are pretty consistent in their diagnosis. But in the middle range (especially the highlighted rows), the diagnosis is all over the place. This is because even a trained doctor’s human eye is not extremely efficient at this image recognition problem. We know that CNNs are great at solving this problem. Google research team used the shown architecture of CNN with 26 layers for this purpose.
The results showed that the algorithm’s performance is on par and even better than the ophthalmologists’ diagnosis.
Reinforcement Learning — Treating HIV
HIV affects about 36 million people worldwide. It requires lifelong treatment with antiretroviral (does not kill the virus but prevents its further growth) drugs. HIV is tricky because the rapidly mutating virus results in drug resistance. This means we have to keep changing the drugs with time to outsmart the virus. This means that the sequence of administered drugs is very important.
Using a simple illustration, if we were to treat a patient with different combination of drugs and if one of drugs makes the patient resistant to other drugs, we will need to think about the sequence of prescribed drugs and paying a close attention to how the drugs affects patient health for sustainable long term care.
From the graph it can be gleaned that future long-term care depends on the current and previous order of drugs even for a similar viral load. Research has shown that humans are not great at keeping track of history of the past medications and figuring out the impact of these drugs on future care by running real-time analytics. This is where Machine Learning can serve as a powerful ally to address the problem. We will be looking at Reinforcement Learning in this context as the suggested architecture for this problem.
Reinforcement learning involves a Deep Neural Network(Agent) trying to navigate an unknown environment to achieve a tangible output with reward feedback at regular intervals. With this constant feedback, the model understands the actions which reward it in a positive or negative way. In our example, the model continues to track numerous biomarkers with every administration of drug (and provides feedback in the form of positive reward or negative warning) and uses this knowledge over time to determine the best strategy for long term care.
Natural Language Processing (NLP) — Unlocking Healthcare Data with Clinical NLP
There is a wealth of information in today’s Patient Records which is not readily available in discrete data columns (because of EHR structure) to be consumed by other application programs or computational units. This insight is embedded deep in physician or specialist notes collected during consultation, clinical narrative reports, radiology, laboratory data, etc. Providers, researchers, and insurance companies use EHR data to find trends in diseases and treatments to find options for better patient care. Machine Learning helps unravel the key information often embedded deep within such text script in these diverse health records. One of the popular architectures used for NLP is called Long Short Term Memory (LSTM) neural network.
In this architecture A represents a full Recurrent Neural Network cell that takes the current input of the sequence (a word in this example), xt, and outputs the current hidden state hi, and passes it to next cell. The inside of a A cell includes a forget gate x, sigmoid and tanh activation layers etc. The detail of this architecture is beyond the scope of this article and will be reviewed in a future article. The key takeaway is that LSTM networks are capable of modeling temporal and sequential aspects of data and are powerful architectures used for processing text in NLP.
Conclusion
Machine Learning and Deep Neural Networks have been used in cutting edge research institutions to find solutions for complex health problems. These solutions have been able to provide long term quality care and save patient lives. We saw a few examples on how Deep Learning algorithms have been used in real world scenarios in the Healthcare industry.
