For decades, improving clinical trial outcomes has focused on reducing time and cost. That focus, and the challenges of a global pandemic, has ignited the digitization of the clinical trial life cycle. Decentralized Clinical Trials (DCTs), which put the patient at the center of the trial experience and incorporate digital technologies, and artificial intelligence (AI) are combining to transform clinical research. Organizations are marrying DCT methodologies with AI in innovative ways to reshape workflows across the clinical lifecycle, from trial design and patient recruitment to evidence generation—and are seeing striking decreases in both the time and cost of clinical studies.

Today, there are at least 25,000 clinical trials offering a form of open source data. With the right analytical tools and AI, the industry can take advantage of this rich data to optimize patient recruitment, increase patient retention, reduce timelines, and maximize study results in future trials. AI can also expedite target validation, trial design, and patient identification.

DCTs can also reduce time and cost of clinical studies by expanding patient access and retention as a result of allowing patients to participate safely and conveniently in their homes. DCTs also minimize geographic, transportation, costs, and other barriers that can improve the diversity of participants. The net effect is an improvement in data quality as site-based transcription is eliminated plus an increase in efficiency for patients, sites, and study teams. Together, AI and DCTs can be a powerful force.

It is time to deploy AI with purpose. To do so, AI engines require more than just mountains of data. To provide accuracy, the data needs to be closely representative of the target population. Traditional clinical trials, as has been highly publicized in recent months, are often comprised of participants from urban areas and are predominantly white. DCTs fuel AI engines since they inherently capture massive real-world, real-time data from diverse patient populations more inclusive of the diverse world in which we live. By combining both innovations, companies will improve three crucial areas of clinical trials: trial design, patient recruitment, and evidence generation.

Customized Trial Design that Drives Results
The application of AI and machine learning can significantly improve trial design in a host of ways—from identifying the ideal sample size to what data should be experimentally generated, to target validation and trial execution. Industrywide efforts to target underlying biological pathways of disease also drive a need for more precise patient screening protocols and trial design.¹ Machine learning, a primary method of AI that can consume vast data sets and utilize complex algorithms to learn from it, finds patterns and makes predictions about future state. To advance the development of these methods, organizations like Machine Learning Ledger Orchestration for Drug Discovery (MELLODY) and Machine Learning for Pharmaceutical Discovery and Synthesis (MLPDS) were formed.²

Vaccines provide a powerful example of AI’s impact on trial design, especially in light of the race for a COVID-19 vaccine. As the vaccines for COVID-19 progress to late stages, AI can help identify signatures of vaccine response. Similar to the adoption of viral load as a surrogate endpoint, which drastically reduced HIV trials from four years to six months, understanding immune signatures of vaccine response can accelerate vaccine trials. Recognizing the critical role of adjuvants in the efficacy of vaccines, a 2019 malaria vaccine study by a group from U.S. Army Medical Research and the Walter Reed Army Institute demonstrated that the combination of machine learning with immune-profiling could most likely identify adjuvant-specific immune responses that can serve to inform decisions on adjuvant selection.³ The method can also enhance vaccine safety by providing valuable insight regarding which immune profiles are associated with reactogenicity.⁴

In addition to improving safety and efficacy, machine learning can also expedite the development of new vaccines, as evidenced by a 2016 collaborative effort of the Georgia Institute of Technology, Emory University, and the Centers for Disease Control and Prevention.⁵ The group utilized DAMIP, a machine learning method, to predict the ability of a vaccine to provide effective immunity within one week of administration. More recently, a group at the University of Michigan employed Vaxign (reverse vaccinology tool) and Vaxign-ML machine learning tool to identify the most appropriate targets for a COVID-19 vaccine.⁶

Not only has machine learning been deployed in other vaccine trials to understand critical success factors for COVID-19 vaccine trials, it has also been employed to identify sites with non-COVID-19 trials that would benefit most from decentralization. For example, Medable’s data science and digital trial design teams have been tracking the spread of COVID-19 to understand the impact on clinical trials worldwide. The heatmap (Figure 1) shows the proximity of active oncology study sites to COVID-19 hotspots in the U.S. Combining these data with the fact that many people remain undiagnosed, reinforces the need to enable the continuation of trials remotely, especially for vulnerable patients.

This data has proven essential to proper planning and COVID mitigation. Machine learning has even predicted future outbreaks, providing key insight to which future trials and sites would be most impacted. Such endeavors directly reduce the barriers, time, and cost of trials. Sponsors are finding considerable value in this innovation. One sponsor commented, “the COVID-19 predictive analytics have enabled us to optimize our feasibility process while ensuring the safety of participants and avoiding delays. The ability to keep pushing forward during a pandemic on behalf of cancer patients everywhere has been invaluable.”

Expediting the Speed and Diversity of Patient Recruitment
Patient recruitment remains one of the largest obstacles in drug development. Broad marketing campaigns through avenues like social media have proven expensive and largely ineffective since it is nearly impossible to identify participants that meet clinical criteria at the right time and in the right place. For rare diseases specifically, it is difficult to meet minimum patient enrollment targets within specified timeframes.

The diversity of trial participants is also a challenge that has generated a significant effort in the industry. Currently, trials do not adequately represent the population with respect to gender, race or ethnicity and many in the industry are initiating efforts to address the problem. For example, Nature published a 2018 study showing that while African Americans comprise 13% of the U.S. population, they only make up 5% of clinical trial participants.⁷

The combination of AI and DCTs is key to overcoming the bias challenge since together; they can advance the identification, enrollment and participation of underrepresented groups by removing barriers such as selection bias, geography, transportation, and more. In addition to ensuring diversity of participants, AI can better identify candidates who are most likely to respond to the intervention and even project which patients are least likely to drop out. Some organizations are using advanced neural networks and Bayesian algorithm powered tools to find those candidates—the needle in the haystack.⁸

Similarly, many interventional studies require a time-sensitive clinical state that can be difficult to forecast and short in duration. For example, there are critical windows in both Alzheimer’s disease and COVID-19 during which patients are suited for participation in trials. It can be nearly impossible for patients, and even sometimes their primary physician, to know when this window is open. AI methods employing probabilistic models can identify patients in advance of the window. Machine learning applied to clinical data, like that maintained in electronic health records (EHRs), can identify regions and individuals with higher pre-recruitment probability for screening success.

A research group at Mount Sinai Hospital in New York applied an unsupervised deep learning method to EHR data of approximately 700,000 patients to determine patient representations that provide enhanced clinical predictions.⁹ The model, dubbed DeepPatient, tested over 76,000 patients, encompassing 78 diseases. It predicted future diseases from cancer to schizophrenia, one of the most challenging diagnoses. DeepPatient held consistently better results than other models. AI and machine learning applications like those leveraged in this project have laid the groundwork to identify research participants at the most critical clinical juncture.

AI allows trial managers to identify patients faster and easier while reducing the rate of screen failures. Once potential participants are identified, the use of DCTs removes geographical and transportation barriers to screening and participation, resulting in improved speed, retention and diversity of recruitment.

Real-Time, Real-World: Enhancing Data Quality in Evidence Generation
Decentralized clinical trials require digital avenues for evidence generation in ways that traditional clinical trials do not. Evaluating the effects of an intervention requires digital endpoints as well as systems that can autonomously interpret physiologic data to detect side effects and monitor patient safety. Therefore, DCTs collect more data that require efficient and accurate analysis and management. AI can discriminate signal from noise and plays a critical role in the development of digital biomarkers.

It is widely known that machine learning-generated biomarkers from preclinical datasets are crucial to drug discovery. In 2017, the Federal Drug Administration approved pembrolizumab for cancers with a certain genetic biomarker.¹⁰ This milestone marked the inaugural tissue agnostic approval of a treatment based on a patient’s biomarker status as opposed to tumor histology. Identification of new preclinical biomarkers can lead to more tissue-agnostic therapies, expediting treatments for multiple disease types.

Organizations are now developing cloud-based machine learning systems to identify new digital biomarkers. For instance, Medable is using machine learning-based informatics to convert time series signals into a set of histogram signatures that can be used with various supervised and semi-supervised learning methods to create biomarkers. These informatics can also be used for patient identification, feasibility, and data source extraction. Data can be ingested from smartphone sensors, surveys, devices, lab data and more. For instance, gait stability has been monitored in cancer patients and detected fatigue and cognition levels. It is highly accurate and reliable but also has a user-friendly interface with data visualization.

New machine learning-based informatics provide users varying levels of expertise with low-level and high-level interfaces. It is not strictly required to be a data scientist in order to perform complex data preparation operations and to create and train highly accurate models. Biomarker identification of extends to a multitude of disease categories. Its ability to monitor patient data in real-time provides has the promise to increase measures of patient safety, ensuring accurate reporting of outcomes in a highly scalable manner.

The value of digital biomarkers stretches beyond indicators of disease to applications in digital therapeutics. Digital therapeutics can shape patient behavior and treat various conditions through digital technologies.¹¹ Using machine learning and other AI approaches with digital therapeutics to monitor and predict data of patient symptomatology can create digital biomarkers,¹¹ resulting in a feedback loop which lends itself to precision medicine.¹² DCTs provide for a much broader application of this method, culminating in voluminous data that makes the AI models even smarter.

In the trial setting, digital biomarkers can predict treatment response from a digital behavioral intervention at the participant level.¹³ The combination of AI and digital therapeutics also enable enhanced real-time clinical measurements without being compromised by the bias of patient recall.¹⁴ In the health care delivery setting, AI and digital therapeutics are essentially a medical device, enabling greater automation and less costly population health management. In addition to monitoring patient symptoms and predicting outcomes, they allow for dynamic adaptation to individual needs, goals, and lifestyles through course correction—particularly important for patients in rural or underserved areas.¹⁵

The Marriage of AI and DCTs Offers Clinical Potential
At no time has the wisdom of Plato been more evident where necessity is the mother of invention, and now adoption. The current pandemic has driven the fastest adoption of technology and innovation in clinical trial history, and the power of AI married with the vast data available through DCTs makes for a fruitful relationship, particularly in trial design, patient recruitment, and evidence generation. Leveraging AI and DCTs has become a fundamental imperative to developing safe therapies as fast as possible. AI’s continued maturation will expand its use across all stages of the patient journey, and ultimately lead to incredible breakthroughs in personalized medicine to sure cancers and rare diseases.¹⁶ 

References

1. Gerlovin, L., Bonet, B., Kambo, I., and Ricciardi, G. The Expanding Role of Artificial Intelligence. Clinical Leader, June 9, 2020. See full resource here.
2. Ibid.
3. Chaudhury, S. Duncan, E.H., Tanmaya, A., Dutta S., Spring M.D., Leitner, W. W., & Bergmann-Leitner, E.S. (2020) Combining immunoprofiling with machine learning to assess the effects of adjuvant formulation on human vaccine-induced immunity, Human Vaccines & Immunotherapeutics, 16:2, 400-411. See full resource here.
4. Ibid.
5. Lee, H.I., Nakaya, F.Y., Querec, T.D., Burel. G., Pietz, F.H., Benecke, B.A., & Pulendran, B. (2016) Machine Learning for Predicting Vaccine Immunogenicity. INFORMS Journal on Applied Analytics. 46:5, 368-390. See full resource here.
6. Ong, E., Wong, M. U., Huffman, A., & He, Y. (2020). COVID-19 coronavirus vaccine design using reverse vaccinology and machine learning. bioRxiv : the preprint server for biology, 2020.03.20.000141. See full resource here.
7. Gerlovin, et al. Expanding Role.
8. Diversifying clinical trials. Nat Med 24, 1779 (2018). https://doi.org/10.1038/s41591-018-0303-4
9. Miotto, R., Li, L., Kidd, B. A., & Dudley, J. T. (2016). Deep Patient: An Unsupervised Representation to Predict the Future of Patients from the Electronic Health Records. Scientific reports, 6, 26094. See full resource here.
10. Boyiadzis, M. M., Kirkwood, J. M., Marshall, J. L., Pritchard, C. C., Azad, N. S., & Gulley, J. L. (2018). Significance and implications of FDA approval of pembrolizumab for biomarker-defined disease. Journal for immunotherapy of cancer, 6(1), 35. See full resource here.
11. Palanica, A., Docktor, M. J., Lieberman, M., & Fossat, Y. (2020). The Need for Artificial Intelligence in Digital Therapeutics. Digital biomarkers, 4(1), 21–25. See full resource here.
12. Guthrie NL, Carpenter J, Edwards KL, et al. (2019). Emergence of digital biomarkers to predict and modify treatment efficacy: machine learning study. BMJ Open, 9:e030710. See full resource here.
13. Palanica, et al. Digital Therapeutics.
14. Guthrie, et al. Digital Biomarkers.
15. Palanica, et al. Digital Therapeutics.
16. Ibid.
17. Vamathevan, J., Clark, D., Czodrowski, P., Dunham, I., Ferran, E., Lee, G., Li, B., Madabhushi, A., Shah, P., Spitzer, M., & Zhao, S. (2019). Applications of machine learning in drug discovery and development. Nature reviews. Drug discovery, 18(6), 463–477. See full resource here.