Original article was published on Artificial Intelligence on Medium
Artificial Intelligence in the Healthcare Industry
How AI can lead the future of diagnosis and drug discovery
Artificial intelligence is one of the technologies that will power the future of healthcare. AI isn’t intended (and won’t be able) to completely automate healthcare processes. AI robot doctors and nurses aren’t the goal here. Rather, healthcare professionals working side by side with AI will usher in a new era of unprecedented efficiency in patient care. In this article, we’ll go over how AI can improve the current systems around patient diagnosis and drug discovery as well as consider the roadblocks preventing AI from being implemented in healthcare.
Currently, misdiagnosis is a huge problem: in America alone, an estimated 12 million people suffer a diagnostic error each year and an estimated 40,000 to 80,000 deaths occur each year in U.S. hospitals related to misdiagnosis (Source: Fierce Healthcare). Through AI, specifically convolutional neural networks (CNNs), diseases can be diagnosed from medical imaging more accurately. Before we go over how CNNs work to diagnose diseases, let’s understand what CNNs are.
Artificial neural networks are modeled after the biological neural networks of the human brain and can make predictions based on patterns recognized in past data. A CNN is a type of neural network that typically includes convolutional layers and max-pooling layers. A convolution applies a filter onto the collection of pixels that make up the input image. This results in an activation. Repeated applications of this filter result in a map of activations called a feature map which essentially tells the computer about an image. Following the convolutional layer is the max-pooling layer. In the max-pooling layer, the filter on the image checks for the greatest pixel value in each section (the size of the section is specified by the programmer) and then uses the maximum pixel values to create a new, smaller image. These smaller images help the computer run the model much faster. When the convolutional layers and max-pooling layers are connected to the input and output layers of a neural network, the model is able to use past labeled data to make predictions of what new images are.
Author’s Note: Explanation of CNNs reused from my previous article “Teaching Computers to See”
These x-ray images above depict the lungs of a normal person, a person with bacterial pneumonia, and a person with viral pneumonia. A dataset with labeled images of each condition would be used to train a CNN. After training the model, we can feed the x-rays of patients to the model and it would classify the x-ray as indicating either a healthy person or a person infected with bacterial pneumonia or viral pneumonia.
Implementing AI for diagnosing conditions seems very promising: studies have found that models have performed on par with human professionals in correctly diagnosing disease. Recently, the Stanford Machine Learning group developed a model that can diagnose pneumonia in a mere 10 seconds! In the future, integrating AI alongside diagnosticians can reduce the chance of a misdiagnosis.
Drug discovery is another area of healthcare that AI seems poised to disrupt. Due to the high level of complexity involved, bringing new drugs to the market is time consuming (> 10 years) and expensive (costs $2.6 billion on average). Furthermore, the likelihood of a drug becoming approved by the FDA is less than 12% (Source: PhRMA). By utilizing neural networks in the search for novel drugs, pharmaceutical companies aim to simultaneously reduce the time and money required in this process.
Neural networks are comprised of an input layer, hidden layer, and output layer. The input layer is where data is fed, the hidden layer is where neurons with weights and bias perform computations, and an activation function in the output layer gives the final output. When neural networks are trained on large, labeled datasets, their predicted outputs are compared to the actual outputs and the error function is used to update the weights/bias (backpropagation). The ability of neural networks to quickly learn and make predictions using large amounts of information makes them ideal for drug discovery.
In the first step of the drug discovery process, researchers seek to understand a human disease at a molecular level. Once researchers formulate an idea, they focus on identifying a drug target that treats or prevents the disease when a drug compound interacts with it. With the drug target identified, researchers screen a multitude of compounds until they pinpoint a few compounds that might eventually become medicine. This process alone can range from three to six years. Neural networks can drastically speed this up; for example, the AI drug discovery company twoXAR has truncated the process to around only three months.
It’s also important to keep in mind that AI’s application to healthcare will face various hurdles.
- It is inevitable that AI will make diagnosis errors in some cases; patients might show increased scrutiny to an AI error as opposed to a human error. Doctor can be sued for misdiagnosis under medical malpractice law but litigation for an AI misdiagnosis does not currently exist.
- The need for large datasets poses a few challenges to the adoption of AI in healthcare. Medical data isn’t readily available which makes it hard to develop effective AI models. Furthermore, collecting data from patients raises privacy concerns. One potential solution to this is storing patient anonymously through a blockchain ledger.
- Bias is another important problem to take note of. Bias in AI is a problem that exists beyond healthcare applications and extends to a widespread concern about the technology overall. Since AI models make predictions based off what they learned in the training dataset, they might not be able to generalize to all patients if the training data is skewed towards patients with a particular race, gender, location etc. Thus, ensuring that a diverse range of populations are represented in the training dataset is crucial.