Author: Nuha Mohammed, Intern at Jeeva Informatics Solutions, 2018; Student at Thomas Jefferson High School for Science and Technology
Medicine has evolved to encompass a great deal of technology. From doctors to dentists to surgeons, medical professionals around the globe are utilizing high performance diagnostic tools and surgical technologies to treat patients. Most importantly, artificial intelligence (AI), inclusive of machine learning (ML) and deep learning (DL), has been impactful in enabling us to diagnose disease much earlier.
At Jeeva, we are implementing AI algorithms in the backend of a robust mobile application and web portal to diagnose rare diseases using an individual’s biological and demographic profile in combination with health data retrieved from cell phones and wearable devices. Patients can enroll in clinical cohort studies and support research as they receive medical input and updates as they participate.
Before we get into the applications, I will provide an overview of AI’s progression and explain the oftenly confused distinctions between AI, ML, and DL.
AI vs ML vs DL
Artificial Intelligence first introduced “intelligent”, algorithm-based computing. Conceptually, we can imagine AI as a combination of if-statements leading to a particular outcome. One step further, and with machine learning we can predict an outcome using our algorithm and improve the accuracy of future outcomes using the present outcome: an obvious reason to call it machine learning. Machine Learning is more dynamic and possesses the ability to improve itself based on previous data given. Finally deep learning further reduces the number of steps in this process and increases efficiency. DL is a subset of ML and it is based on neural networks. It is named for functioning using deep neural networks (with 3 or more hidden layers). Today, when we refer to AI, we are referring to a wide spectrum of computing inclusive of basic AI, ML, and DL.
Image credit: Nvidia via Towards Data Science– Differences between AI, Ml, and DL
AI at a Basic Level
In a basic AI model, we train the algorithm by pointing input data to a specific output. Multiple variations of that data are grouped. This is called input to output mapping and it is an essential concept in all of AI. Fundamentally, the greater the amount of data in each set, the more precise the algorithm.
Image credit: XenoStack via TheStartup – Machine Learning and Deep Learning processes.
Now, let’s get into some of the trending applications of machine learning in diagnostic medicine!
- Medical Imaging/Computer Vision
Images of the eye can tell you a lot about brain function. Many diagnostic AI tools analyze the retinal blood vessels from plain images taken of the eye. For example, diabetic retinopathy, can be diagnosed using retinal images because of distinct features displayed in the retinal nerves such as cotton wool spots and neovascularization, both of which can be detected by AI.
Another interesting application of AI in medical imaging is automatic detection of leukemia using an AI. With a manual approach, microscopic cell images are tediously searched by scientists to look for characteristic properties. Changes in blood composition, including white blood cell count, are strong indicators of Leukemia. Why not train an AI to look for these characteristic properties so we can reach the diagnosis in seconds? Researchers at Thapar University in Patiala India, used an AI to detect leukocytes and count white blood cells. They also extracted physical features from the images and quantified these features for final image classification.
MRI brain scans and x-rays are also important visual data sources for AI. Specific abnormalities in neuron shapes and quantification of tumor vessels help diagnose brain diseases, including Glioblastoma, using MRI results.
Image credit: gfycat – How a Neural Network Learns
Above, you can see how a neural network learns to detect cavitary lesions in a TB patient’s lungs. The input consists of x-ray images.
- Numerical Lab Data
Lab tests include A1C levels, creatinine levels, liver function values, ammonia, iron, and a myriad of other values. This wealth of numerical data generated from just one blood or urine test is then accumulated for long-term (cohort) studies serving to identify patterns/characteristic features associated with diseases.
- Sensor Data
Wearable technologies such as the Fitbit and Apple Watch contain a variety of sensors, inclusive of the accelerometer and gyroscope. Your cellphone itself contains these two sensors. These two specifically can be used to monitor the gait of an individual. In addition, because of these sensors being integrated into the cellphone, hand tremors can be detected and analyzed along with gait. With tremor frequency data from multiple Parkinson’s patients, an AI could be employed for pattern recognition to identify tremor frequencies associated with Parkinsonian tremors.
- Voice Data
Speech patterns are also indicative of some neurodegenerative diseases such as Parkinson’s, of which slurred speech is a common symptom. AI can analyze similarities between speech graphs through the presence of specific peeks. Essentially, when a new voice pattern is recorded, it is tested with the AI classifier for similar patterns of slurred speech.
Image credit: Medium – Sampling a sound wave
By now, you should have identified a recurring process in :
- Identify symptoms of the disease you are studying.
- Correlate to health metrics: Can this symptom be displayed in an image? If so, what technology would be used to produce this image (i.e. MRI, cellphone, microscope, PET scan, CT scan)? Or is this symptom better quantified with numerical data?
- Aggregate Data: If this data is open-source, then great! If you can’t get a hold on any related data, contacting organizations such as the NIH, and also requesting clinical trials would be a feasible option. At Jeeva, we are creating mobile applications to serve as versatile patient recruitment, engagement, and data sharing platforms for clinical trials, one step further than the traditional clinical trial method.
- Select the appropriate approach: If using the DL approach, structures such as an Artificial Neural Network (ANN) or Deep Neural Network (DNN) would be used. A few structures specific to ML include Support Vector Machines, Neural Networks, kNN classifiers, K-Means, and Random Forest. etc., More specifically, selecting the algorithm depends on the type of data inputted and outcome expected.
- Code! There are many resources such as tensorflow, Keras, Weka, and H2O.ai which make AI way easier. Once you have identified the best structure/algorithm to use, select the right AI platform/library for you based on your experience level. If you are a beginner, you should look at a few tutorials and practice on sample data sets before you get started on your project. Good luck!
From the researcher’s perspective, large amounts of data are required for data analysis procedures and creating AIs. This data, if not found online from previous trials, needs to be acquired through clinical trials. Patient recruitment is very difficult for in-person trials. However, many new mobile application technologies have eased the patient recruitment process. As a part of my internship at Jeeva, I am working on developing a mobile application to serve as versatile patient recruitment, engagement, and data sharing platform.
Further, the app will be leveraging phenome and genome data of patients of rare, undiagnosed diseases. The data will be analyzed with a machine learning approach to identify patterns and identifiers associated with the patients’ symptoms.
Enrolling in these studies, enables you to get personalized treatment recommendations and health suggestions along the way, as you contribute to the progress of biomarker and treatment knowledge for rare, undiagnosed diseases!
Through the extension and application of biological research, developers are creating computational algorithms for diagnosis. AI offers many integral applications in diagnostic medicine which are unparalleled to its applications in other fields. Using machine learning, we might even be able to diagnose patients suffering from rare, undiagnosed cases. AI will only continue to grow and enable us to reach new milestones in medical technology!