How to launch a AI Clinical Digital Assistant? (CDA)

Objective: To build a new feature for clinical digital assistant (CDA) which will personalize the user experience for the doctors.

There can be various goals for the system in this case:

1) Record and Recommend Medications for the patient based on history
2) Summarize visits
3) Enable transcriptions
4) Suggest treatment paths/options based on Acuity, type of visit, goals of patient, etc
5) Make patient visits more personal and desirable (although no patient visit is desirable from a patient’s point of view) and free up doctors time to enable more interaction with the patient (rather than spending time on computers)
6) Minimizing Clinician’s errors, optimize doses, build trust, establish guidelines

Context: Let’s say we have access to data from a leader in EHR applications for clinical workflows. Doctors use it everyday to input Patient information, Transcriptions, medications and other data.

Assumptions:
1) For the sake of discussion, we will focus on primary care doctors (user type) It’s possible to segment primary care doctors based on region (city, town etc), demographic data of cities (various factors such as no. of people in various age ranges, percentage of people with chronic diseases, Rate of no. of healthcare incidents per location etc)

I believe primary care doctors is right choice to begin with because most patient journeys start here, and AI can make a lot of difference in patients life by uncovering right data and insights that can augment a primary care physician(although this assumption needs to be tested by
some user research)

2) We are not building an assistant for patients directly, we never want to replace doctors/clinicians

3) Most of the assistant experience will be conversational, both delivered using an app as well as desktop (inside their workflow/ omnichannel). So most of the user experience will be chatbot based, including tools to capture images (including handwritten notes), transcribe voice, as well as to get input data (wherever required)

4) We have access to the right data sources (in this case most importantly EHR records, that are segmented based on patient demographics, as well as clinical data (location, type, size, facilities,speciality, etc).

5) Availability of Data Sources: Multiple types of data might be available (Notes (text), Images of Scans, Electronic recordings from medical devices, Vitals (text and Numerical data, tables, graphs), Lab results, etc)

6) Real time data from clinics gets added to EHR records for patients, allowing us to perform data analysis across EHR records? (Data can be anonymized in this case for meeting HIPPA regulations, but this needs further study to make sure their is compliance on multiple levels regarding data handling) (We will talk about this more in a detailed data governance project)

7) Also assuming the project is feasible, we have the right resources to execute, given the problem statement is valuable enough, with a clear path to adoption and revenue.

Datapoint:
Primary care physicians use it to automate mundane tasks, for e.g. before visits summary (based on history), track medications, transcriptions, etc but in no particular observable fashion

It seems as if not every patient visit has the same goal, sometimes it’s a routine checkup,sometimes critical, sometimes it’s more lab related, etc). It’s hard to generalize the “next bestaction/task” for the doctors. We need more analytics to inform the next best suggestion actionm for the doctors (Part of data strategy).

Focus during User research
1) Identify and rank the tasks that a doctor needs to perform
2) Try to establish minimum set of tasks that must be completed using AI/automation (make
sure this aligns with their feedback and our observations) (for e.g. notes,Vitals, treatment
path etc)
3) Understand the tasks in detail, from a doctors point of view

A typical UserFlow in this use case
1) Assistant has access to the doctors scheduling system, so is able to remind the doctor about the upcoming patient’s visit, and summarize the plan of action.

2) Patient arrives, doctor records vitals, patient describes current state, doctor uses voice transcription to record patient data or by simply having an “ambient mode” where CDA listens to the conversation between patient and the doctor.

3) Doctors can constantly provide feedback to the assistant by providing (microfeedback, using clicks, saying yes or no etc) so that the assistant can improve suggestions (medications, treatment plans, possible diagnosis, etc). This will also ensure AI models align with goals defined by the stakeholders.

4) Doctors end the session, triggering a few backend features that summarize visits, perform analytics, identify last minute suggestions for patients, set up the next appointment, and send the patient the plan of action.

5) Execute backend admin related tasks (billing, medical coding etc)

DATA Sources to be used in building the CDA

1) EHR records (text, images, tables, distributions, structured text)
2) Publicly available clinical research data (nih.gov, pubmed etc)
3) Social data (condition related anecdotal data on social forums, video testimonial transcriptions etc) (this can be a differentiator)
4) Patient and doctor descriptions/conversations/inputs (in some cases, if these can be integrated into the clinical workflows) (to provide more contextual data for prompt engineering later)
5) Data from other apps (for e.g. Apple health, Calm and other wellness related apps, by forging unique partnerships with them)
6) Data in various languages (for translation)

AI models that need to be built and fine tuned for this project
1) Summarization models (to summarize visits)
2) NLP Models for Voice transcription (and translation in various languages)
3) Large language models (fine tuned to healthcare data sets, to provide QuestionAnswering, Classification, Natural Language understanding, Intent detection and reasoning)
4) Models that can uncover big contextual data and generate insights
5) Helper models for fact checking (for e.g. to fact check doctors prescription/ or suggestion for a new treatment plan, identify risks for the patient etc)
6) Helper models for monitoring and improving accuracy of base models (this is still an area of research)
7) Models that can extract data from tables, perform analysis across data sets to uncover patterns to provide predictions.
8) Recommender model that suggests NBA (Next Best Action) to the doctors (based on various factors such as visit type, acuity, patient history, existing research on the topic , effectiveness of treatment for certain demographics etc)

Overall expected Challenges during the project:
1) Delivering Value quickly where there are 10s of possible solutions, ways forward.
Incrementally delivering the value by providing a seamless experience for primary care physicians will be challenging.
2) Challenges in driving the accuracy of transcriptions to nearly 100%
3) Fact checking AI generated insights for the doctors
4) Aligning AI with system and user goals, as well as business goals.

5) Managing customer expectations around AI
6) Constantly updating and expanding data sets
7) Achieving almost near 100% accuracy for the patients well being.
8) Risk of Sensitive data leaks

Things to consider in our Data Strategy:
1) Evolving set of data sources to be identified (from various diverse sources to enable cross checking, fairness, consistency etc)
2) Define the goals of data strategy, define users, how they will access data, etc (be data ready)
3) Create data pipelines tailored for specific use cases and scenarios (tie this to your data ingestion plan)
4) Find and prioritize data that’s contextual and annotated (to ensure high quality)
5) Create analytics component to understand the data better, uncover patterns, as well as allowing us to shape the data according to the goals that we want to achieve with the AI
6) Create a data governance strategy (that encapsulates healthcare
regulations/compliance requirements around handling patient data)
7) Collect and analyze user generated data, use it to shape your existing data sources
8) Build Data redaction and Data quality check tools wherever necessary

The MVP considerations:
So as discussed above, let’s say we want to build an assistant focused on just primary care physicians. We have also done the required user research and prioritized the problem for now.

In addition to the “must-have” administrative tasks that are common for various primary care physicians, let’s say we will add one additional feature that will utilize our AI system.

1) Let’s say we want to focus on delivering the feature called “NBA” or Next Best Action for the primary care physician based on the historical and real time EHR data. This feature is a good choice because it is a high-value feature that will enable patient well being, save doctors vital time trying to uncover insights and relationships between the Vitals (data).
For example, based on Vital signs recorded on a patient visit , CDA can suggest a treatment plan or lab test, as the next best Action for the patient.

This is assuming that CDA can understand Vitals and provide suggestions based on historical data or because of established healthcare metrics around vitals (such as normal blood pressure, heart rate etc )

2) We will build the required conversation flow for the above mentioned tasks (for thedoctors) (Make sure user experience is flawless)
3) Provide a way to collect feedback after each suggested NBA by the CDA (to add humans in the loop, to ensure our system is learning). Doctors should be able to provide quick as well as detailed feedback on the type of output they are looking for (we can use this shape our AI model, data sources, and inference process)
4) Metrics used to test the success of our feature launch will be: Accuracy of NBA (as pointed out by doctors as well as fact checking, consistency of NBA’s across various patient visits, time saved for doctors etc)