Artificial intelligence and other forms of advanced data analysis frequently are viewed as completely impartial and free of bias. However, when the people that set up the processes inadvertently introduce bias, the results can be impacted.
Rich Christie, chief medical officer of AiCure, recently connected with Outsourcing-Pharma to talk about the problems that bias in AI can lead to in trials and development.
OSP: Could you please share some perspective on bias around clinical trial data? Specifically, how can AI and other advanced analytical tools (largely seen as objective) be slanted toward one group and/or against another?
RC: Bias can arise in AI when the data used in initial training sets aren’t completely representative of the diversity present in subsequent patient populations. After all, the strength and generalizability of AI models are reflective of the underlying training data sets, which is why consistently evaluating the quality and quantity of data sets is critical.
For example, the data sets that many computer vision tools trained on in the past lacked variability and diversity along with a number of dimensions, typically because many of them were based on computer programmer volunteers, a group historically limited in diversity. This means that when visual tools are used with patients of diverse backgrounds, the inherent bias can not only impact research operations, but it can also affect patient outcomes and ultimately impact the ability to contribute meaningful data to studying the safety and efficacy of drugs.
The consequences of biased AI can range depending on how the tool is deployed, either in clinical development or real-world settings. In one scenario, an AI system aimed at identifying optimal investigator sites might sub-optimally skew sites chosen to participate in recruiting and managing participants in a clinical trial. In another scenario, a study could fail if a trial is using insufficient AI to determine patient engagement and enrollment rates – an incorrect number of early terminations within a previously unidentified sub-population could inadvertently cause a slant in results.
In another example, if computer vision technologies in a clinical trial don’t work uniformly on patients with diverse skin tones, there could be inadvertent bias baked into study results that causes a misperception of patient benefit across the entire population.
In 2022, we will need to work together to ensure the technology our patients and pharmaceutical companies use has the foundation it needs to foster equality and reach its potential. By focusing on the diversity of training data sets, formal generalizability of results, and real-world confirmation, companies in the pharmaceutical space can develop stronger algorithms to optimize drug development and patient outcomes.
OSP: How is AI helping drive progress in precision medicine?
RC: AI-powered predictive analytics can help drive precision medicine in numerous ways, including guiding timely, personalized interventions that help patients stay on track with their treatment, and providing sponsors with valuable insights into patient and site behaviors both during trials and before trials even begin to stratify risk.
The more we can capture the patient experience directly and the nuance of the behavior at the level of individual participants, the better we may be able to match a therapeutic to a specific patient’s needs. AI is continuing to take on an important role in revealing these insights. Specifically, video and audio-based digital biomarkers can capture a patient's response to treatment particularly well due to their consistency, automation, and ability to scale across geographies and patient populations.
Quality of life assessments such as a patient’s ability to button their shirt or sign their name can be conducted to really understand how patients are responding to treatment and its impact on their quality of life. Instead of relying solely on in-person clinical assessments or patient self-reported data, video and audio digital biomarkers can measure subtle changes in a patient’s behavior that otherwise might be missed.
Additionally, AI can help predict ahead of time a patient’s ability to adhere to a trials’ protocols based on their previous behavior, which can help sites focus on patients who may have trouble staying on track and tailor helpful guidance towards them. Sponsors can also use this data to optimize patient pools by understanding who may be better positioned to contribute quality data to answer the questions the study is asking.
OSP: What do you feel might be the biggest trends to look out for in 2022? Please feel free to talk about potential obstacles, as well as areas of opportunity.
RC: Looking ahead to 2022, one area slated for growth is the use of open-source data platforms to advance AI. Video and audio digital biomarkers are an untapped resource for accurately measuring patient behavior and improving a clinical trial’s objectivity, but currently, the proprietary nature of digital biomarker algorithms keeps researchers from exercising them to their fullest potential, hindering their clinical validation and refinement.
As an industry, we need to continue encouraging open-source platforms so the scientific community can gain access to algorithms, jointly contribute to their advancement, and further validate digital biomarkers as a legitimate means to understanding disease.
Secondly, we will continue to see a renewed focus on mitigating the impact of bias on AI in healthcare. The role of AI in healthcare is only growing, and AI-powered insights will continue to affect every element of drug development, from clinical research to commercialization to populational assessments of health.
With the increased role of AI comes the possibility for it to perpetuate unseen biases if the industry is not proactive in its management. If AI is unbiased and reliable, it has the power to help sponsors address real-world variability, better understand how medications work and their impact on specific patients, and bridge clinical findings into an increasingly diverse environment of real-world data collection.
OSP: What does AiCure have planned for the coming year that you might want to preview?
RC: AiCure’s AI platform is evolving to deliver the predictive and unbiased insights needed to optimize drug development throughout the drug life cycle. Currently, much of the data used for decision-making in clinical trials are lagging indicators of participant behavior, and responses are typically reactive rather than proactive, leading to wasted resources, longer trial timelines, and delayed care.
To solve this problem, AiCure is working to enable structured data collection through secure, compliant handling of unique PHI data so users can understand how patient behavior drives value and can access predictive insights to enable intervention when it matters. The AiCure platform offers a single place to collect, aggregate, annotate, build and test novel assessments and algorithms, providing customers with the ability to develop symptom-specific AI models that allow a more accurate measurement of signs of disease through audio and video data.