Why using AI in clinical trials becomes the new normal

Studies show that clinical trials of new drugs last nine years on average and cost around $1.3 billion to carry out. The cost of failed clinical trials, meanwhile, ranges between $800 million and $1.4 billion. And the fact that 90% of all drugs end up failing clinical trials only complicates the matter. In traditional clinical trials, doctors and researchers manually look for participants, and patients have to be physically present to enroll and undergo evaluation. The treatment also occurs on site through scheduled visits. This remains a safe approach to developing new remedies. However, it is slow and lacks the flexibility required to compose complex therapies and address the needs of smaller population segments that are often heterogeneous. Additionally, this approach doesn’t have the capacity to integrate and process data from hospitals, research centers, private practices, and patients' homes. Researchers would struggle with participant recruitment, and would request patients to visit trial sites for systematic condition reviews and monitoring, which could increase the chances of patient dropout. Artificial intelligence and its subtypes can help resolve these issues.

AI can integrate data from multiple sources, including electronic health records (EHRs), research papers, past clinical trials information, and special medical case studies. It can also handle the continuous stream of data from personal medical devices. AI-driven clinical trial technology can aggregate, clean, process, manage, and visualize all this information in a way that helps clinicians understand a given disease and the potential that different chemical compounds offer in countering it. While predictive analytics in healthcare helps foresee how patients can react to the proposed remedies. Gaining access to insights derived from all this information in a timely manner will empower researchers to make more informed decisions fast. Here is how AI can benefit different aspects of clinical trials.

Artificial intelligence has many benefits in the healthcare sector. For example, since the pandemic hit, pharmaceutics extensively used AI to speed up clinical trials of potential COVID-19 vaccine candidates. There are five major applications of AI in clinical trials. The technology:

Research shows that poor clinical trial design can prevent a potentially efficacious drug from demonstrating efficacy, wasting all the resources spent on developing this medication. But designing clinical studies is challenging as pharmaceutical companies need to look through vast amounts of data, 80% of which is unstructured and hard to analyze. AI for clinical trials can help aggregate and process all this data and find useful patterns. For example, it can derive the right regulatory protocols, strategies, and patient enrollment models that suit the country of the trial. AI can also help identify the best timing for conducting the study. This will result in encountering fewer protocol amendments, patient dropouts, and regulatory violations. The Tufts Center for the Study of Drug Development found that one substantial protocol amendment can prolong a trial for three months and cost between $140,000 and $530,000 depending on the trial’s phase.

There are three main patient-related issues that hinder clinical trials.

Traditionally, patients can hear about relevant trials from their physician or search a corresponding database, like the national US registry of clinical studies. These sources are not sufficient, as doctors are not aware of all the ongoing trials and patients might find scrolling over governmental websites overwhelming, especially given their recent diagnosis. Enhancing clinical trials with AI allows for sifting through patient data, such as EHR and medical imaging, to compare patient characteristics to the study’s eligibility criteria to identify the right individuals for this particular trial. AI is powerful enough to select a homogeneous set of participants, which is challenging with the conventional methods. An AI startup Paradigm uses its vast database of oncology studies to recruit patients for trials. The startup can match people newly diagnosed with cancer and speed up their enrollment in trials. While 23andMe, a personal genetics company based in California, suggests clinical studies to its clients based on their genetic makeup.

Research shows that approximately 30% of participants tend to quit clinical trials. This results in increased expenditure and time needed to complete the study. Recruiting one patient for a clinical trial costs on average $6,500, while replacing a patient when the trial is already underway costs even more. We can resolve both of these issues with a rigorous patient selection. As mentioned in the previous point above, AI investigates patient data and can look beyond the study’s admission criteria, minimizing future dropout.

Candidate participants need to go through evaluations to ensure they meet the inclusion criteria, which demands their physical presence. And depending on their location and job flexibility, they might not be able to visit the trial’s facilities in the dedicated time. AI can streamline wearable technology deployment, allowing patients to take some evaluations at home. Then machine learning algorithms can aggregate and analyze the data. For example, a medical startup TytoCare offers connected examination tools and underlying mobile apps that enable patients to capture measurements from their lungs, heart, skin, throat, etc. and send it to clinicians.

AI can analyze data on available doctors, patients, and climate conditions at different geographical locations and visualize it on a map, which helps pharma companies select an investigator site with the biggest potential. One example of using artificial intelligence in site selection comes from Innoplexus. This clinical trials AI company helps pharmaceutical firms design and prepare for studies with its Clinical Trial Comparator technology. It offers dashboards for visualizing information that helps prioritize sites for prospective clinical studies, including proximity to competitor clinical trials, geography, and candidate population. Innoplexus also developed a customized AI-powered dashboard with filters that allows its clients to integrate third-party data and set thresholds and metrics for their own site selection criteria.

Medication non-adherence is rather common. Studies indicate that 50% of Americans fail to take their long-term chronic medication as instructed. And according to the World Health Organization, medication adherence can have an even bigger impact than the treatment itself. In clinical trials, the process of manually tracking medication adherence is prone to error, as it relies on patients’ memory. And doctors often use unreliable recording systems, such as pen and paper, which can lead to information loss. Deploying wearables together with clinical trial AI allows researchers to monitor patients’ actions through automated data capturing instead of waiting for the patients’ manual reports. For instance, AiCure, one of the prominent AI clinical trial companies, developed an interactive medical assistant that can spot patients at risks of non-adherence. This technology also allows patients to take a video of themselves swallowing a pill as a proof that they actually did it. The assistant can identify the right patient and the pill, confirming adherence to the responsible doctor.

Clinical trials consume and output massive amounts of data. Every participant would generate excessive information, such as adherence data, vital signs, and any other intermediate feedback. AI can aggregate, analyze, and present it to clinicians in a readable format. Also, with the help of medical IoT devices and the Internet of Bodies, clinicians can monitor patients in their home in real time. This means processing large amounts of data daily. AI can take over this task and spot and report any deterioration in patients’ condition, ensuring patient well-being and minimizing dropouts. Another interesting benefit is that machine learning algorithms can identify patient cohorts within a trail that merit further investigation. For instance, if the trial doesn’t seem to yield the expected results, AI can identify participants with specific conditions that seem to benefit from the investigated drug or treatment for sub-trials.

Despite the efforts put into unifying medical data, there are still multiple healthcare IT standards, and health data interoperability is still a challenge. This makes it hard to integrate patient information from medical organizations that use different EHR software. Not to mention that some doctors still rely on handwritten notes. Even though AI’s operations are hindered by lack of interoperability, the technology can also help overcome this problem. Natural language processing (NLP)-based models can extract clinical data, such as symptoms and diagnosis from diverse heterogeneous sources, and aggregate this information into the trials database instead of normalizing health records and other sources. One example is Deep 6 AI, which uses NLP to parse diverse EHR systems. The company was valued at $140 million in its latest fundraise. However, the job of NLP algorithms is not that straightforward as there is no unified terminology that doctors use to express the same concept. For instance, some physicians refer to a heart attack as “myocardial infarction” or “myocardial infarct,” while some just jot down “MI.” Therefore, clinical trial AI models need to be equipped to recognize all these variations.

AI has its specific difficulties that it brings to every field where it is applied. If you want to discover more about AI, check out our recent article on AI implementation challenges and how much AI costs. Here are two of the most relevant challenges artificial intelligence brings to clinical trials:

At the moment, there is still no reliable, fully-automated replacement for the manual data annotation process required to train artificial intelligence models used in clinical trials. This task is time-consuming, and the results are often tailored to individual healthcare providers or specific diseases. “Right now, there is no such thing as an NLP engine that takes any clinical notes written from any physician and can understand what the notes say,” said Noemie Elhadad, a Biomedical Informaticist at Columbia University, emphasizing the limited reusability of trained NLP models.

AI can develop bias if the training dataset is not representative of the actual population, as the generalizability of the model depends on the diversity that it saw during training. For example, improperly trained models can skew site suggestions for clinical trials or can perform poorly on patients with darker skin tones. Even algorithms that are well-trained can acquire bias as they continue to learn on the job. Therefore, it is important to conduct timely independent audits to catch on any inappropriate behavior and eliminate it. “AI is a living medical product that needs to be constantly tweaked and recalibrated,” says Dr Leo Anthony Celi, Principal Research Scientist at Massachusetts Institute of Technology. He believes that AI and machine learning in clinical trials need to be viewed as a separate product, independent of the medical devices the technology is used with. Therefore, AI-powered solutions have to be assessed independently and frequently.

Accenture predicts three waves of improvement in traditional clinical trials, some of them will take a long time to mature.