Understanding the Impact: FDA's Take on AI and ML in Medical Devices

The rapid advancement of technology in healthcare has led to the integration of Artificial Intelligence (AI), particularly in the subset known as Machine Learning (ML), into a growing number of connected medical devices. 

AI holds immense potential in the medical device industry, offering numerous benefits such as enhanced data management, remote surgery capabilities, diagnostic and procedural assistance, and support for clinical trials. 

Additionally, AI can optimize medical device manufacturing processes, leveraging machine learning to improve efficiency and mitigate risks. By ingesting vast amounts of data and learning from errors, computers equipped with AI algorithms streamline operations and minimize the potential for human error. 

The integration of Artificial Intelligence in connected medical devices for performing or supporting medical applications comes with the need for adherence to new regulations outlined by entities like the Medical Device Regulation (MDR), In Vitro Medical Device Regulation (IVDR), and the U.S. Food and Drug Administration (FDA). These regulations stipulate that AI-driven medical devices must meet state-of-the-art requirements, providing concrete evidence for repeatability and reliability.

How Are Artificial Intelligence and Machine Learning Transforming Medical Devices?

One of its most significant advantages is AI/ML's potential to derive valuable insights from the vast daily volumes of healthcare data. In many aspects of our health and daily routines, digital health technologies are becoming increasingly influential, with AI/ML driving essential advancements. The FDA's commitment to ensuring the safety and efficacy of these devices while enabling them to realize their full potential in enhancing people's well-being remains a focal point of its public health mission.

There has been a notable surge in interest in connected medical devices incorporating AI/ML functionality in recent years. For instance, connected blood pressure monitors, such as sphygmomanometers, replicate the actions of trained clinicians by detecting and reporting Korotkoff sounds indicative of systolic and diastolic blood pressures. Portable defibrillators assess heart waveforms to determine the need for defibrillation and can autonomously administer the required treatment.

These technologies, often supplementing or even replacing direct clinician involvement, have extended the reach of healthcare beyond traditional facilities, reaching homes, workplaces, and areas with almost nil or absent trained clinicians. However, it's crucial to note that these devices don't operate independently of human reasoning; instead, they employ pre-validated clinical protocols to diagnose medical conditions or deliver therapy.

Radiology leads the way in FDA-approved medical devices and algorithms, particularly with AI/ML-based global image-reading software solutions. For instance, Arterys Inc. offers three algorithms—Arterys Cardio DL, Arterys Oncology DL, and Arterys MICA—integrated with leading Picture Archiving and Communication Systems from vendors like Siemens Healthineers AG and GE Healthcare. Six of these algorithms are oncology-focused, with three dedicated to mammography analysis and three for CT-based lesion detection.

These algorithms are integrated with leading PACS from vendors like Siemens Healthineers AG and GE Healthcare. PACS is a technology used in medical imaging to securely store and digitally transmit electronic images and clinically relevant reports. The integration allows seamless collaboration between AI algorithms and existing infrastructure, making it easier for healthcare professionals to incorporate these advanced tools into their workflows.

Arterys Cardio DL is designed for cardiovascular imaging analysis. It uses AI to analyze medical images, such as MRI scans, to detect patterns and abnormalities related to the cardiovascular system. This includes identifying cardiac conditions, measuring blood flow, and assessing the heart's health. The AI assists cardiologists & radiologists in making more accurate and efficient diagnoses.

Arterys Oncology DL focuses on oncology-related imaging. The algorithm is trained to analyze medical images, particularly those related to cancer detection and characterization. This includes identifying tumors, assessing their size and characteristics, and monitoring changes over time. The AI aids radiologists and oncologists in early detection, treatment planning, and ongoing monitoring of cancer patients.

Arterys MICA stands for Medical Imaging Cloud AI. It is a platform that leverages AI for various medical imaging applications. This platform likely supports deploying various algorithms for different medical imaging modalities and specialties, providing a flexible and scalable solution for healthcare providers.

 AI algorithms can analyze large datasets quickly, potentially reducing the time required for image interpretation and diagnosis. AI can assist in detecting subtle patterns or abnormalities in medical images that may be challenging for human eyes, leading to more accurate diagnoses. Especially in oncology, AI can aid in the early detection of tumors, allowing for timely intervention and improved patient outcomes.

Additionally, there are two algorithms for brain image analysis, specializing in stroke and hemorrhage detection, and six for enhancing image processing, including noise reduction and radiation dosage reduction. Furthermore, four algorithms cater to acute care, addressing pneumothorax assessment, wrist fracture diagnosis, and triage of head, spine, and chest injuries using systems like Aidoc Medical BriefCase. Finally, two algorithms focus on cardiovascular assessments, specifically heart ejection fraction evaluation. 

Brain Image Analysis Algorithms (Stroke and Hemorrhage Detection) are designed to analyze brain images, mainly focusing on stroke and hemorrhage detection. AI is used to identify subtle signs of these conditions in medical images such as CT or MRI scans. The integration of AI enhances the speed and accuracy of diagnosis, facilitating prompt medical intervention in cases of stroke or hemorrhage.

Image Processing Algorithms (Noise Reduction and Radiation Dosage Reduction) are AI-based image processing algorithms that aim to improve the quality of medical images. This can include reducing image noise, enhancing clarity, and optimizing radiation dosage. By leveraging AI, these algorithms produce higher-quality images, essential for accurate diagnosis while potentially minimizing radiation exposure to patients.

Acute Care Algorithms (Pneumothorax Assessment, Wrist Fracture Diagnosis, Triage of Head, Spine, and Chest Injuries) are AI algorithms for acute care that are tailored to assess specific conditions quickly and accurately. For example, in pneumothorax assessment, AI can rapidly identify the presence of a collapsed lung in chest X-rays. Wrist fracture diagnosis algorithms analyze images to assist in identifying fractures in the wrist. Triage algorithms for head, spine, and chest injuries prioritize cases based on severity, aiding healthcare providers in efficient decision-making during emergency situations.

Cardiovascular Assessment Algorithms (Heart Ejection Fraction Evaluation) are for cardiovascular assessment, such as evaluating heart ejection fraction using AI to analyze cardiac imaging data. Ejection fraction is a crucial metric for assessing heart function. AI can assist in accurately measuring this parameter, aiding cardiologists in diagnosing and managing cardiovascular conditions.

AI algorithms can analyze medical images rapidly, potentially reducing the time between image acquisition and diagnosis. They help to detect subtle abnormalities that might be overlooked or misinterpreted reducing variability between different healthcare providers and ensuring consistent diagnostic quality.

In acute care settings, AI algorithms assist in quickly identifying critical conditions, allowing for timely interventions and improved patient outcomes. Also, Image processing algorithms contribute to reducing radiation dosage and promoting patient safety by minimizing exposure to ionizing radiation during medical imaging procedures.

Cardiology has witnessed significant progress, leading to the development and FDA approval of four groundbreaking medical devices and algorithms. The majority of investments are directed towards advancements in detecting cardiac rhythm abnormalities. Notably, the AI-ECG Platform and Eko Analysis Software have received FDA approval, alongside two algorithms, EchoMD AEF software and EchoGo Core, which also have applications in the field of Radiology.

The AI-ECG Platform likely involves the application of artificial intelligence to electrocardiogram (ECG) data. AI algorithms can analyze ECG signals to detect patterns and abnormalities in cardiac rhythm. This technology enables more efficient and accurate identification of cardiac conditions, such as arrhythmias or heart electrical activity abnormalities.

Eko Analysis Software is designed to work with electronic stethoscopes and likely utilizes AI to analyze heart sounds. AI algorithms can help identify subtle variations or abnormalities in heart sounds that may indicate certain cardiac conditions. This can aid healthcare professionals in diagnosing heart-related issues during routine examinations.

EchoMD AEF (Atrial Ejection Fraction) software likely focuses on analyzing echocardiograms, explicitly assessing the function of the atria in the heart. AI can be applied to these imaging data to provide more accurate measurements, identify abnormalities, and evaluate cardiac function. This technology can be particularly relevant in diagnosing and managing conditions like atrial fibrillation.

EchoGo Core is another algorithm designed for echocardiogram analysis. Echocardiograms use ultrasound to create images of the heart, and AI algorithms can be employed to analyze these images. EchoGo Core likely automates the measurement of key parameters, such as ejection fraction and chamber dimensions, contributing to a more efficient and standardized assessment of cardiac function.

AI algorithms can identify subtle signs of cardiac abnormalities, enabling early detection and intervention. Automation of analysis tasks by AI can significantly reduce the time required for interpreting cardiac data, allowing healthcare professionals to focus on decision-making and patient care.

In response to diabetes's widespread impact, considerable efforts have been devoted to innovating blood glucose management. Pioneering solutions include the Guardian Connect System by Medtronic and the DreaMed Diabetes system from DreaMed Diabetes Ltd. Furthermore, the Internal Medicine sector has witnessed the integration of AI/ML-based interpretation through the FerriSmart Analysis System by Resonance Health Analysis Service Pty Ltd, specifically designed for assessing liver iron concentration.

The Guardian Connect System likely incorporates AI algorithms to analyze continuous glucose monitoring (CGM) data. CGM devices continuously measure glucose levels throughout the day and night. AI interprets this data, recognizing patterns, trends, and potential glucose-related events.

The DreaMed Diabetes system is likely designed to optimize insulin therapy. It employs AI algorithms to analyze various factors, including insulin dosages, blood glucose levels, carbohydrate intake, and physical activity data. The AI aims to provide personalized insulin recommendations.

The FerriSmart Analysis System likely utilizes AI algorithms to assess liver iron concentration. Liver iron concentration is crucial in managing conditions like iron overload, which can be associated with various diseases, including certain types of anemia.

Over the past decade, the FDA has been actively reviewing and authorizing an increasing number of devices utilizing AI/ML technologies across various medical fields. This authorization often occurs through 510(k) clearance, De Novo requests, or premarket approval. The FDA anticipates this trend to persist in the coming years. 

What Statistics Say About AI/ML-enabled Medical Devices

The annual growth of AI/ML-enabled devices slowed to 15% in 2021 and 14% in 2022, due to the FDA regulation, following a substantial 39% increase in 2020 compared to 2019. Projections for 2023 suggest a significant resurgence with an expected increase of over 30%.

In 2022, 87% of authorized devices were in Radiology (122), Cardiovascular at 7% (10), and various other specialties at 1% each. As of July 2023, 79% of devices authorized in 2023 are in Radiology (85), followed by Cardiovascular at 9% (10). 

Between August 2022 and July 2023, most AI/ML devices that received clearance underwent the FDA's 510(k) pathway, with only two devices obtaining de novo clearance during this period. Source: www.fda.gov

From August 2022 to July 2023, an analysis revealed that GE HealthCare and Siemens Healthineers took the lead in the number of cleared AI/ML devices. GE HealthCare has the highest total of authorized AI/ML devices.

Radiology consistently leads in the number of submissions and the steady growth of AI/ML-enabled devices within specialties. Machine Learning models vary in complexity, ranging from shallow to deep learning. The trend leans towards hybrid approaches, combining different algorithms for feature generation and classification to ensure the safety and effectiveness of devices.

As of January 2024, the current list of AI/ML devices on the FDA’s website includes 692 entries.

[Also read: Top 5 Trends That Are Shaping the Future of Advanced Medical Devices]

Overview of the New FDA Guidelines Related to AI and ML in Medical Devices 

The draft guidance outlines the FDA's current perspective on change control plans for connected medical device software, explicitly emphasizing the role of a predetermined change control plan (PCCP) in premarket submissions. Including a PCCP in a submission serves the purpose of delineating software modifications that necessitate a new marketing submission without such a plan.

According to the FDA, a PCCP functions as a valuable tool to facilitate iterative improvements in medical device software while ensuring a reasonable assurance of device safety and effectiveness. The three essential components of a PCCP are as follows.

PCCP Component 1: Description of Modifications

This section precisely identifies modifications that, contingent on FDA approval of the PCCP, can be implemented without requiring a new marketing submission. These modifications, aimed at maintaining or enhancing device safety and effectiveness, must undergo verification, validation, and documentation in adherence to the manufacturer's quality system.

PCCP Component 2: Modification Protocol

The modification protocol outlines the procedures to be followed when developing, validating, and implementing the modifications described in Component 1. This protocol includes four crucial components for each proposed modification:

• Data management practices
• Re-training practices
• Performance evaluation
• Update procedures

These components enable the FDA to assess the proposed modifications, ensuring compliance with good Machine Learning practices and determining whether the protocols reasonably assure safety and effectiveness.

PCCP Component 3: Impact Assessment

The impact assessment involves a comprehensive benefit/risk analysis for each specified modification, including any associated risk mitigation measures. This analysis compares the current version to the proposed changes, evaluating the potential interactions among all proposed modifications.

While manufacturers can employ the PCCP to plan various device modifications, the guidance document underscores that a new marketing submission is necessary when a modification affects the device's intended use or significantly impacts safety or effectiveness. This approach ensures that crucial modifications undergo thorough scrutiny and regulatory evaluation to uphold patient safety and device effectiveness.

Future of Artificial Intelligence  and Machine Learning in Medical Devices

Medical device companies seeking to implement AI/Machine Learning changes in their products now have a structured pathway to follow. The FDA recommends utilizing the Q-Submission process to obtain feedback on a proposed predetermined change control plan (PCCP), especially for combination products and high-risk devices such as life-sustaining, life-supporting, or implantable devices.

With the recent AI/Machine Learning in Medical guidance, developers can incorporate a PCCP into their initial application, allowing them to seek authorization for future updates. The FDA advocates for early communication with the relevant review division overseeing the proposed PCCP to streamline this process.

For future submissions involving AI/Machine Learning in medicine, it is crucial to note that the guidance specifies that in a determination of substantial equivalence, particularly when the predicate device was authorized with a PCCP, the subject device must be evaluated in comparison to the version of the predicate device that was previously approved or cleared before any updates or modifications were made under the PCCP. This approach ensures a thorough assessment while considering the evolution of the device over time and maintaining a focus on patient safety and efficacy.

How BioT Can Help 

We at BioT have a team of engineers and regulatory experts with extensive experience in the MedTech industry. We are prepared to guide you in successfully integrating AI/Machine Learning in medicine and bringing your FDA-approved AI/Machine Learning medical device to market. We recently helped Neeteera to deploy their proprietary AI on BioT’s Distributed Medical Device Platform to improve patient care. To know more or know how we can help, get in touch with us today.