AI-driven patient retention and engagement for clinical trials

Introduction

Bringing a new drug to market is a slow and laborious process. According to Tufts Centre for the Study of Drug Development, it takes on average between 7 to 10 years and costs more than $2.5 billion to bring a new drug to market. (1) More than half of this time and capital is spent on clinical trials. Despite that, the success rates of clinical trials are low. According to a report, only around 12% of studies between January 2000 and April 2019 ended in success. (2) Drug-development programs fail for a host of reasons, ranging from flawed study design and failure to recruit enough patients to a high participant drop-out rate. (3) An average 30% of patients drop out during the course of a clinical trial and 40% are non-adherent after 5 months, which represents a huge opportunity cost for the industry.(4)

To address some of the challenges organizations are currently facing in clinical research, they are turning to artificial intelligence (AI) to reduce drug development cost, improve the trial designs, enhance patient recruitment and retention efforts, and eventually speed up the drug development process.

In this blog, we discuss how life sciences customers can implement AI-driven patient retention & engagement solutions using AWS technology. Additionally, we provide an architecture blueprint for implementation of such solutions.

Using Amazon ML services, clinical researchers can develop a AI-driven patient retention and engagement solution, without having to worry about the undifferentiated heavy lifting involved in setting up an AI/ML solution, hence achieving their patient retention and participation goals. Patient experience can be enhanced in the following ways:

Facilitating access to information and providing an improved participation experience through chatbots, voice assistants, intelligent search, and intelligent contact center solutions
Perform remote follow-ups through an intelligent contact center solution, which not only reduces burden on the trial participants but also enable researchers to capture variety of data without the patient needing to be onsite
Using mobile apps to capture behavioral, study, and therapeutic data
Using the mobile app, wearables, sensors, and data from other systems to develop a patient profile and identify non-compliance, predict drop outs, and adverse events
Virtual clinical trials: Leveraging remote endpoints captured through mobile (ePROs), wearables, sensors and other remote data capture technologies to better understand trial participants’ behavior during the study.
Analyzing the conversations to provide feedback to clinicians/site coordinators to improve the process/procedure for patient engagement

Improving patient retention and engagement through AWS technologies

In today’s interconnected world, majority of people have developed a preference for easy and rapid access to information online. Given the complex nature of the drug development process and the demand it imposes on study participants, it is crucial that they have timely and easy access to all relevant information. Instead of having to scour through pages of study content for answer to a simple question, patients could get a direct response from an intelligent search engine such as Amazon Kendra using a chat bot interface. Instant access to relevant and accurate information will not only improve trial performance in the long-run but will enhance the participants’ empowerment, experience, and their desire to continue participating in the trial.

Chatbots and voice assistance (e.g. Amazon Echo and Alexa) can also form part of an omnichannel patient engagement strategy. In addition to mobile & web applications, these channels can be used to develop conversational applications that facilitate clinical trial data collection (e.g. patient surveys, vitals measurements), verify completion of care pathway tasks and allow patients to report health concerns. The conversational applications can also prove useful in ensuring patient adherence through the use of chat or voice reminders. Finally, voice assistants can act as a trial companion by offering answers to a patient’s questions or concerns regarding the trial and treatments involved.

The tasks assigned to study participants can be overwhelming and they might often struggle to follow the recommended schedule. Forgetting to follow study guidelines (e.g. taking medications) is cited as a reason for non adherence among 50-60% of patients.(5) Using the patient’s medical and behavioral data (obtained through a mobile app, medical devices, EMR, questionnaires etc.), we can develop a patient’s profile. This data, can be used by an algorithm to determine each patient’s individual preferences, habits and schedules, within the limits of privacy and regulatory frameworks. An AI application can subsequently send notifications and guidance to help patients follow the study guidelines. Additionally, a visual confirmation can be obtained using a solution like AWS DeepLens to ensure adherence. Facial recognition and object detection can be used to ensure the patient is taking the right medication and the right dosage.

An intelligent contact center solution, powered by Amazon Connect, can be used to provide patient insights in real time, which can help researchers and trial administrators better understand and respond to patient needs and improve the overall patient experience. AWS artificial intelligence (AI) services, Amazon Transcribe, Amazon Translate, and Amazon Comprehend, can respectively be used to enable transcription, translation, and analysis of patient interactions. The intelligence gathered from such data can be used to have more informed conversations with the patient and better understand patient motivation and participation. The data captured from patient interactions can also be used to detect adverse events and take necessary action. Amazon Comprehend Medical, which enables mapping of

Reference architecture

The above use cases provide a range of possibilities for leveraging AI and data-driven solutions to improve patients’ participation in clinical trials. As technology evolves and organizations adopt novel approaches to manage clinical trials, technology will play an increasingly crucial role. The below diagrams illustrate the typical architecture using Amazon ML services to facilitate patient engagement in a clinical trial setting. The focus is on facilitating data capture through mobile, sensors, and voice channels, facilitating communication and easy access to information, and use of Amazon ML services to capture additional valuable information.

The architecture is composed of four work-streams, namely:

1) Device data capture

The proposed architecture for capturing data from patient’s mobile and other devices is outlined below:

A patient logs into a mobile device and enters data on a regular basis or data is collected passively through wearables, sensors, activity trackers, and other smart devices
The device(s) stream the data to AWS IoT Core, which writes the data to Amazon Kinesis Data Firehose in real-time.
Kinesis Data Firehose writes the data to an Amazon S3 bucket.
Data collected in the S3 bucket is processed using AWS Glue, a managed extract, transform, and load (ETL) service, to integrate with other data sources and provide a consolidated view of study participants.
The integrated dataset is stored in an S3 bucket.
The integrated dataset is used by trained machine learning models to derive inference (e.g. detect non-adherence).
If the inference results require a notification to the patient, the notification is sent using Amazon SNS.

A key part of the above architecture is the provisioning of medical devices on AWS IoT. AWS IoT Core offers a feature called Fleet Provisioning, which is designed to help customers easily onboard large numbers of manufactured devices, from consumer devices to industrial equipment. We have a reference implementation of Fleet Provisioning available on GitHub, which can be used as a baseline to build the above proposed solution. Refer to the aws-iot-fleet-provisioning GibHub repo for a more hands-on experience of provisioning devices and capturing device data using AWS IoT Core.

2) Voice Assistant and chatbot interactions

As discussed before, voice assistants and chatbots can be used to improve patient interaction and collect data. The below architecture shows how this can be achieved using AWS services:

The above architecture demonstrates patient interaction via two different channels, namely an Amazon Alexa enabled voice assistant (e.g. Amazon Echo) and a chatbot powered by Amazon Lex.

The patient interaction with Amazon Alexa involves the following steps:

1a. The patient starts an interaction with an Alexa enabled device
2a. Data is streamed to Alexa Voice Service , which makes use of Alexa Skills Kit to handle a range of patient requests. This includes, for example, a request educational content about the trial, participate in a survey, or schedule an appointment.
3a. Alexa Skills Kit SDK triggers the relevant AWS Lambda function depending on the patient request.
4. If the interaction involves patient request content concerning the study or questions frequently raised by patients, the request is redirected to Amazon Kendra, an intelligent search service, which retrieves the relevant information from an S3 bucket.
5. If the interaction involved submission of new data from the patient, the data is stored in a database (e.g. Amazon DynamoDB)
6. If the interaction involves integration with a third-party system such as a Clinical Trial Management System (CTMS), a Lambda function is used to integrate with such systems.
7. Newly arriving data stored in the database is processed using AWS Glue, and integrate with other data sources (if deemed appropriate).

The patient interaction with a chatbot would involve the following steps:
1b. Patient logs into a mobile or web application by submitting their security credentials.
2b. Patient’s security credentials are verified using Amazon Cognito.
3b. Patient starts interacting with the chatbot powered by Amazon Lex and makes a request. The requests made by the patient will be similar to the ones they can raise through the voice assistant leading to steps 4 – 7 as described above.

Our previous published blog on “Building Clinically Validated Conversational Agents to Address Novel Coronavirus”, provides an example implementation for setting up conversational bots using Amazon Lex with a DynamoDB database. Additionally, our blog on “Integrating Amazon Kendra and Amazon Lex using a search intent” provides a useful reference implementation for integrating the two services.

3) Intelligent contact center

The below diagram illustrates how an AI powered contact center solution can facilitate patient engagement.

When a patient calls into the Amazon Connect call center, their call progresses through a contact flow and customer audio is streamed in real time through Amazon Kinesis Video Streams.
An AWS Lambda function consumes the audio stream and uses the Amazon Transcribe service to convert the audio into text
Once the AWS Lambda function receives the results from Amazon Transcribe, it stores the text data in Amazon DynamoDB.
The solution also leverages Amazon Translate and Amazon Comprehend to get translated (if applicable) and annotated text data.
The annotated and translated text is integrated with other data sources using AWS Glue.
The integrated dataset is stored in a central S3 bucket. Investigators and data scientists can mine the contact trace records for additional insights to improve the overall patient experience and draw additional insights such as patient sentiment at a certain stage of the trial.

Our documentation on AI Powered Speech Analytics for Amazon Connect provides a reference implementation for the above solution.

4) Data integration and analytics

Once the data is centralized from all the different channels that patient interacts with, the researchers and data scientists involved in the study can perform analysis of the data using Analytics tools available in AWS. The below architecture provides showcases a typical architecture for such use cases:

Trial investigators access the data for analysis and reporting using Amazon QuickSight and Amazon Athena
Amazon Athena presents the data stored in S3 bucket(s) in the form of tables and views based on table definitions defined in AWS Glue Data Catalog
1. Glue Data Catalog is periodically refreshed using a Glue Crawler
Results for queries executed by users of Amazon QuickSight and Amazon Athena is retrieved from the S3 bucket
Data Scientists use Amazon SageMaker Notebook instances or SageMaker Studio to develop, train and deploy their models
Data is retrieved from the central S3 bucket during model training and evaluation.
Developed and trained models are deployed as endpoints in SageMaker
SageMaker endpoints are used by Lambda functions and other APIs to trigger notifications to patients, research team and other stakeholders via Amazon SNS and AWS IoT

In order to set up the data integration and analytics work stream, you can leverage our quick start for launching data analytics platforms, which found on GitHub here.

Data security, data privacy and other considerations

While use of technology can significantly enhance patient experience and overcome obstacles involved a traditional model, it also poses additional concerns in terms of patient privacy and data security. The increased monitoring resulting from the use of intelligent voice devices, sensors, and wearables means that not only it is important to clearly inform patients about the scale of data collection but also to obtain and record consent. One way this could be achieved is by providing access to information concerning data collection and management through intelligent voice and chatbot interaction channels, which can also be used to respond to patients’ questions or concerns in relation to the subject.

Data security is an equally important concerns for organizations. Services such as AWS IAM (Identify and Access Management), AWS STS (Security Token Service), AWS KMS (Key Management Service), and AWS CloudTrail provide the necessary capabilities for securing your solution in the cloud, including fine-grained access control, data security at rest and in-transit, and logging and auditing. Additionally, there are guidelines available as part of the documentation for securing individual services mentioned in this blog post, which are listed below:

It is also important to note that customers maintain the ownership of data and are responsible for ensuring the security of the data being kept in AWS. The AWS Shared Responsibility Model provides more details about what responsibilities customer must assume to secure their data.

Conclusion

As technological advancements pave the way for an accelerated innovation, they offer significant potential in terms of their ability to enhance clinical trials. From AI-driven protocol development to use of ML for efficient patient recruitment, emerging technologies are starting to influence all aspects of clinical trials. There has never been a better time to explore novel ways of using technology in clinical research, given the challenges posed by the COVID-19 pandemic, a significant drop in clinical trial participation being one of them. In recent years, there has been an uptake in interest towards virtual clinical trials with organizations exploring how technology can play a role in clinical trial execution without compromising the quality of research. Adoption of virtual trials is likely to accelerate in the years to come. Services offered by AWS offer an opportunity for organizations to build secure, scalable, and cost-effective solutions that are customizable to the specific requirement of a clinical trial, without having to dabble with the complexity of developing a new system through a traditional model.

AWS for Industries