This blog explores medical device artificial intelligence (AI) with an overview of  AI, medical applications and devices, the investment landscape, and medical device artificial intelligence regulatory implications.

The term artificial intelligence (AI) has been over marketed and abused.  For instance, a rice cooker with AI, a watch powered by AI, a restaurant ordering app with AI, an air-conditioning unit controlled by AI, a coffee maker run by AI, and more.  Practically any product that is controlled by a computer can be made “cool” by adding “AI” as a suffix.  In my opinion, most of the products on the market labeled as such are using the term as a marketing gimmick, and do not actually use AI.

In reality, what is artificial intelligence? It is a means of equipping a machine to take on tasks or make decisions in a way that mimic and simulate human intelligence and behavior. Simulating human intelligence and behavior is a difficult thing. It requires lots of data, computational real estate and time to train the machine to display human-like behavior.

A related technique is “machine learning”. This is a part of AI that focuses on algorithms that allow machines to learn and change when exposed to new data without being reprogrammed. Going further, another technique called “deep learning” refers to algorithms that allow a machine to autonomously mimic human thought patterns through artificial neural networks composed of cascading layers of information.

To recap, artificial intelligence is a general term describing a process by which a machine can be equipped with human-like intelligence and behavior, and the other two terminologies refer to methods used to make this happen. Keeping it simple, AI can be thought of as an “intelligent system” mimicking human decision-making ability and behavior.

Tools and Techniques

We now have an idea of what artificial intelligence is. The next question is how one can equip a machine to make a decision that simulates human behavior.  There are several tools and techniques used to do this including support vector machines, neural networks, logistic regressions, discriminant analysis, random forest, linear regression, Naïve Bayes Classifiers, nearest neighbor searches, decision trees, Hidden Markov models and others. Support vector machines and neural networks are the techniques used most often, in combination with others.

As an example, I was involved in a project for detecting molecules in the air using a type of odor sensing. We developed a device with 240 sensors that could be used to detect the odor of roasted coffee. For every test, we recorded 240 readings 10 times for a total of 2400 readings. We subjected the device to an additional six varieties of coffee coming from different samples. An algorithm was created to analyze the result using a combination of the techniques mentioned above. The results were probabilistic, identifying the type of coffee with over 90% certainty.

As for the use of AI in medical devices, let us take lung cancer as an example. Developing an early detection tool for lung cancer is very challenging, with most cases being detected at later stage. Using AI as an early detection tool has a strong probability of being a game changer. Imagine having millions of lung CT scans classified into different stages, ethnicities, types of work, exposures to pollution, and other contributory parameters and conditions, along with an AI engine that can analyze these scans in real-time. According to a study conducted in 2019, a deep learning algorithm achieved state-of-the-art lung cancer detection performance of 94.4%. Using 6,716 cases, the AI outperformed radiologists by 11% in false positives and 5 % in false negatives.

Medical Device Artificial Intelligence Applications

In the medical device space, there are many ways AI can be used in a device or system. Here are  8 applications:

1. Diagnosis of heart diseases

A machine learning algorithm (myocardial-ischemic-injury-index) incorporating age and sex paired with high-sensitivity cardiac troponin I concentrations was used to train an AI platform utilizing data from 3013 patients. The platform was then tested on 7998 patients with suspected myocardial infarction and was found to outperform physicians with a sensitivity of 82.5% and a specificity of 92.2%.

2. Detecting Cancer in Mammography

Breast cancer screening via mammography is a widely accepted tool for breast cancer screening and another area where AI can be applied.  With current imaging and analysis tools, cancer cells are often obscured by dense breast tissues. Their appearance on a mammogram can be subtle and can be missed through human error. With the combination of new imaging technologies and an AI engine using a huge set of historical images, the current screening method can be improved by faster analysis, real-time diagnosis, and the absence of human error.

3. Diagnosis of Degenerative Brain Diseases

The diagnosis of neurological conditions such as epilepsy, Alzheimer’s disease, and strokes is a difficult challenge. Current diagnostic technologies (e.g. magnetic resonance imaging, electroencephalogram) produce huge quantities of data for detection, monitoring and treatment of neurological diseases. Analysis of the data tends to be difficult. The use of an intelligent system that accumulate, manage, analyze and automatically detect the neurological abnormalities is crucial. Application of AI in this area will improve the consistency of diagnosis and increase the success of treatment.

4. Detecting a Retinopathy

Diabetic retinopathy (DR) is one of the leading causes of preventable blindness globally. In a study published by American Academy of Opthalmology, a total of 75,137 publicly available fundus images from diabetic patients were used to train and test an artificial intelligence engine to differentiate healthy fundi from those with DR.  The results showed an impressive 94% and 98% sensitivity and specificity, respectively.

5. Cell Sorting and Recognizing Cell Types

A study published in Nature demonstrated the use of a neural network in cell sorting.  The results showed that the system takes less than a few milliseconds to classify cells and provide a decision to a cell sorter to separate individual target cells in real-time. This study shows the applicability of the AI in classifying white blood cells and epithelial cancer cells with 95.71% sensitivity and 95.74% specificity, label-free.

6. Medical Imaging of Liver

AI is gaining popularity in image-recognition applications. AI using deep learning algorithms can automatically make a quantitative and higher efficiency assessment of the characteristics of complex medical images. One application is in imaging the liver to screen for possible liver diseases using radiology, ultrasound, and nuclear medicine. In the image analysis, AI was used in detecting and evaluating focal liver lesions, facilitating treatment, and predicting the appropriate treatment response.

7. In Vitro Diagnostic Tools

AI can be applied to in vitro diagnostics using real-time imaging to capture fluorescence signals as cells pass through a microfluidic channel.  An AI algorithm could be used to differentiate cells by size, shape, and emission wavelength, and can categorize the cells as predictors of  certain diseases.  Moreover, used in combination with other hardware technologies, this can be done in real-time while maintaining the accuracy of the results.  Integrating AI into an in vitro diagnostic platform can improve the performance of the device and diagnostic accuracy.

8. Biosensors for Monitoring Vital Signs

Biosensor-based devices generate huge data sets. Using AI could predict the trends and the probability of disease occurrence.  The integration of AI in cardiac monitoring-based biosensors for point of care (POC) diagnostics are a great example.  Machine-learning algorithms are used with microchip-based cardiac biosensors for real-time health monitoring and to provide accurate clinical decision in a timely manner.

Medical Device Artificial Intelligence Investment landscape

Investing in medical device technology is extremely risky considering it is in a highly regulated space. However, the incorporation of AI in medical platforms has created new perspectives in investment.  This market report  shows that AI in the healthcare market size is projected to grow to USD 31.3 billion by 2025 at a CAGR of 41.5%. This is likely one of the reasons why AI has attracted the attention of huge number of investors as shown in the examples below.

Here are 5 companies that successfully integrated an AI engine in their medical device artificial intelligence products.

Heartflow

HeartFlow, a US company, has raised a total of $577.7MM in funding over 9 investment rounds. Their latest funding was raised on Jun 13, 2019. The company uses deep learning along with 3D models of a patient’s heart and coronary arteries based on imagery from a computed tomography angiography. The company’s computer algorithms solve millions of complex equations to simulate and assess coronary blood flow.  The Heart Flow Analysis has been used to diagnose 40,000 patients.

FREENOME

Freenome, another US company, has raised a total of $507.6MM in funding over 6 investment rounds. Their latest funding was raised on August 26, 2020. The company developed a platform to detect key biological signals from a routine blood draw. The platform integrates assays for cell-free DNA, methylation, and proteins with advanced computational biology and machine learning techniques to understand additive signatures for early cancer detection.  They claim that their technology incorporates a multidimensional view of both tumor- and immune-derived signatures instead of relying only on tumor-derived markers. Their AI pattern recognition technology enables doctors to detect cancer much earlier, and with greater sensitivity and specificity.

ORCAM

OrCam Technologies, a company in Israel, has raised a total of $86.4 MM in funding. They claim to have developed the most advanced wearable AI assistive technology for individuals who are blind, visually impaired, or have reading difficulties.  The product is a small device with a smart camera that attaches to any eyeglass frame, enabling the user to recognize objects, locations, read text, provide directions and more. The device provides these functions by audibly conveying the visual information in real-time.  It can even be used offline.

Paige

Paige is a US based startup that uses artificial intelligence to help doctors diagnose cancer. Paige stands for “Pathology AI Guidance Engine”. The company has raised a total of $195MM in funding over 4 investment rounds, with the last funding raised on January 14, 2021. Paige focuses on breast, prostate and other major cancers. Their product is an AI infrastructure that can operationalize data and algorithms on a large scale. It includes a series of modules for rapid diagnostic stratification, cancer detection, tumor segmentation, prediction of treatment response and overall survival. ​

Sight Diagnostics

Sight Diagnostics, a company from Israel, has raised a total of $123.8 MM in funding. Their latest funding was raised on August 5, 2020. The company aims to bring affordable, scalable and accurate blood diagnostics to the point-of-care. The diagnostic platform includes a cartridge and a reader equipped with an artificial intelligence-driven engine for blood analysis and infectious disease diagnostics based on proprietary machine-vision technology.

Medical Device Artificial Intelligence Regulatory Requirements

The emergence of Artificial Intelligence (AI) is a challenging new front in medical device regulation. As far as I know, there is no regulatory guidance that specifically regulates the use of artificial intelligence in medical devices. One of the major issues with AI is that when using a probabilistic algorithm, it is difficult to explain with certainty how the resulting decisions are made. If it cannot be fully explained or justified, it is difficult for regulators to conclude whether the device is effective and safe to use. Since AI utilizes such a huge set of data, there is no doubt that, in principle, its performance could be far superior to current medical devices.

In April 2019, FDA released a discussion paper entitled “Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD)” that describes the FDA’s foundation for a potential approach to premarket review for artificial intelligence and machine learning-driven software modifications.  As response to stakeholders feedback on the above document, FDA released an action plan  in January 2021 entitled  “Artificial Intelligence and Machine Learning (AI/ML) Software as a Medical Device Action Plan”.

Another important resource entitled “ Algorithms as Medical Devices” was published by the PHG foundation, University of Cambridge, UK.   The document recommended that medical devices must meet existing regulatory requirements.  In Canada, the Canadian Institute of Health Research (CIHR)  in collaboration with Health Canada hosted a discussion on the topic “Introduction of Artificial Intelligence and Machine Learning in Medical Devices”.

This discussion served as a basis to define regulatory requirements and to direct future guidance for medical device manufacturers employing AI and machine learning. If you are interested in learning more about the regulatory developments, I suggest reviewing those documents.

On the other hand, the US Food and Drug Administration (FDA) has granted 501(k) clearance for a number of AI-based system; Such as QuantX, an AI diagnostic tool, aiming to improve breast cancer diagnosis; the Apple watch ECG, which uses electrodes to capture heart rhythm irregularities and cleared as an over-the-counter ECG monitoring device; AIDOC, an AI based system that allows radiologists to identify acute intracranial hemorrhages in head CT scans specifically for the detection of large-vessel occlusions (LVO); and  Zebra Medical Vision, secured five (5) FDA clearances for their AI imaging products.