With traditional approaches to clinical trial data management being manual, cumbersome, and prone to human error, streamlining processes offers huge efficiency- and quality-boosting potential.
Risk-based quality management (RBQM) is speeding up studies while providing greater oversight, but artificial intelligence (AI) can take these gains even further. That’s why CluePoints is constantly working on new ways to utilize the technology in our everyday work.
Laura Trotta, Vice President of Research at CluePoints, is here to answer your burning questions on leveraging AI to improve the quality of clinical trials.
What Is AI?
Simply put, AI is intelligence as demonstrated by machines. While computer intelligence is an aspirational goal, machine learning is the field of study that gives the computers the ability to learn without being programmed. Machine learning includes a wide range of supervised and unsupervised algorithms that allow the computer to learn from data and perform a series of tasks. As part of those techniques, deep learning is a set of advanced machine learning algorithms relying on deep neural networks.
How Are Clinical Trials Suited to AI Solutions?
There is a growing, industry-wide interest in applying ML and DL to the large volume of clinical data collected during a trial. Clinical and operational data is gathered through multiple data sources during study conduct, including eCRF, Labs, ECG, ePRO/ECOA, wearables, and CTMS. The typical Phase III protocol now collects more than 3.4 million data points.1
As both the volume and complexity of this data rapidly increases, advanced ML algorithms can be used to extract more insights. Deep learning algorithms are particularly suited for high volumes of complex data and have been shown to be particularly powerful in dealing with complex data such as text or images. Besides dealing with high volumes of complex data, deep learning techniques are powerful in dealing with decision-making processes that rely on complex cognitive processes. This is particularly relevant to clinical trials where many processes rely on the experience of subject matter experts (SMEs) and their clinical knowledge of the study, indication, and therapeutic area of the drug under investigation.
How Can AI Improve Quality in Clinical Trials?
Many RBQM platforms rely on advanced statistical and ML algorithms. Data quality assessment, key risk indicators, and quality tolerance limits, for example, can all be used to monitor data during study conduct. They work by looking at the study data from all sites in near real time as it accumulates and flagging outlying data points that could signify potential issues.
CluePoints has been working on a series of ML solutions, all designed to improve study efficiency and data quality even further, including a new medical coding module. Adverse events and concomitant drugs recorded in case report forms must be manually mapped to MedDRA or WHODrug dictionaries, which takes time. The DL model guides researchers to the correct corresponding term in seconds, and with more than 90% accuracy.
DL can also be used to improve risk detection. Currently, RBQM analysis will raise a “risk signal” when a potential data issue is identified. These signals, which are to monitor and track any resulting investigations and corrective actions, end up containing realms of free text which is entered by the users as they document their findings.
A root-cause decision feature is a natural language processing algorithm screens all the documentation attached to risk signals and flags those that either lack the required documentation or for which the root cause selected by the user is unreliable. This helps build a consistent database of documented issues that can be used to predict which signals are more likely to represent a real issue at the point of identification. It should allow study teams to prioritize signal review and ensure effective follow-up and findings documentation.
How Does Deep Learning Coding Compare to Auto-Coding?
A deep learning solution is more flexible and can more easily adapt to new entries than auto-coding functions that have a lower hit rate and can only be improved with the manual entry of new terms. While some current auto-coding tools achieve a hit rate of only 60%, a DL solution can reach more than 90%. Medical coding tools should be designed with an API that is able to connect with multiple systems.
To What Standards Are AI Solutions Built?
In 2021, the US Food and Drug Administration (FDA) published Good Machine Learning Practice to promote the safe, effective, and high-quality development of AI- and ML-enabled medical devices. While CluePoints software is not a medical device, we strongly rely on the document’s ten guiding principles when developing our solutions.
AI solutions should leverage a multidisciplinary team of SMEs from across the field in development and train and test its models on independent datasets. Solutions can be tested in conditions similar to those in their intended production use, using studies from various therapeutic areas to avoid any bias in the predictions.
The team should also conduct proof of values testing in which SMEs have direct access to the model’s functions and predictions and work closely with product teams to ensure production monitoring continues after solution release.
What Languages Are Required for AI or ML Project Implementation?
To develop an AI/ML solution, you will need to have machine learning experts with good programming skills, ideally in Python as PyTorch and TensorFlow include Python in their main ML libraries. To benefit from a machine learning algorithm (e.g., by relying on its predictions), you would not need any programming skills.
Still curious about the benefits of leveraging this technology to support your clinical research?
Become familiar with our risk-based quality management solutions and get in touch with the team for more information.
References:
- Getz K, Smith Z, Kravet M. Protocol Design and Performance Benchmarks by Phase and by Oncology and Rare Disease Subgroups. Ther Innov Regul Sci. 2023 Jan.
