This article is part one of a four-part series analyzing a new report on artificial intelligence (AI). Part one broadly defines AI and its building blocks, as well as establishes and classifies potential global use cases.
Artificial intelligence (AI) has been making its way into the healthcare sector, posing a variety of possibilities in disease diagnosis, treatment, and prevention. A recent report, “Artificial Intelligence in Global Health: Defining a Collective Path Forward,” evaluated the progression and the current state of AI in regards to health. The report was a collaboration between The Rockefeller Foundation and United States Agency for International Development’s (USAID) Center for Innovation and Impact, in conjunction with the Bill and Melinda Gates Foundation.
For the purposes of the report, researchers define AI as “the use of computers for automated decision-making to perform tasks that normally require human intelligence”; they further describe machine learning as “a subset of AI that uses algorithms that give computers the ability to learn without being explicitly programmed.”
The researchers evaluated instances in high-income countries (HICs) as well as low- and middle-income countries (LMICs) where AI is being used, tested, or considered in the healthcare sector, leaving them with more than 240 examples.
Building Blocks of AI
The report describes what the researchers consider the fundamental building blocks of AI, which are “technologies that are commonly understood as examples of AI.” There are three categories of building blocks: data, processing, and action.
Data includes, per the authors:
- Computer vision (automated methods used to conduct image-based inspection and analysis)
- Speech recognition (computerized identification and response to sounds produced in human speech)
- Natural language processing (processing and analysis of large amounts of data written in natural language [e.g., narrative])
Processing is three-pronged, they continue:
- Information processing (in AI) (processing of digitized data in ways parallel to human brain functions)
- Machine learning (pattern recognition that learns and improves from experience without being programmed)
- Planning & exploring agents (use of AI for strategies or action sequences by agents, robots, or unmanned vehicles)
Finally, the action consists of four key parts:
- Image generation (automated creation of images using AI)
- Speech generation (automated generation of human-like speech using AI)
- Handling and control (automatic handling of objects using AI methods)
- Navigating and movement (autonomous movement and navigation informed by AI)
Collectively, the examples they determined most relevant to global health are:
- Computer vision
- Speech recognition
- Natural language processing
- Information processing
- Machine learning
- Speech generation
AI Use Cases: Better Together
AI use cases to stratify into four broad functional areas: Population Health, Individual Health (including care routing and care services), Health Systems, and Pharma. Within these four categories were 27 use cases for AI in global health. These groupings are as follows:
- Population Health
- Surveillance and prediction
- Population risk management
- Intervention selection
- Intervention targeting
- Health Systems
- Medical records
- Capacity planning and personnel management
- Claims processing
- Fraud prevention
- Quality assurance and training
- Coding and billing
- Pharma & Medtech
- Clinical trial support and recruitment
- Drug discovery and MedTech R&D
- Drug safety and pharmacovigilance
- Supply chain and planning optimization
- Process optimization
- Real world evidence and HEOR
The researchers further stratified the use of cases under the Individual Health umbrella into two subcategories: Care Routing and Care Services. Care Routing pertains to self-referral, triage, and personalized outreach.
Care Services was specified even further, including a timeline of how use cases are applied: Prevention, Diagnosis, Acute Treatment, and Chronic Treatment and Follow-up:
- Prevention
- Behavior change (exercise, diet, wellness, education)
- Diagnosis
- Data-driven diagnosis (symptom-based, lab-based)
- Image-based diagnosis (radiology, pathology)
- Acute Treatment
- Clinical decision support (treatment guidance, medication prescribing)
- Chronic Treatment and Follow-up
- Medication adherence
- Rehab compliance
- Dietary compliance
Three use cases fall under the Acute Treatment and Chronic Treatment and Follow-up umbrellas: Monitoring (inpatient monitoring, device monitoring), AI-facilitated care (self-care guidance, psych counseling), and AI-facilitated care (robotic surgery, robotic PT).
After these 27 categories were established, the researchers evaluated each use case for impact and feasibility. The impact was valued based on how the use case could increase healthcare access, quality, and efficiency, while feasibility was based on the technology’s current activity and maturity, how it is already being tested in an LMIC context, and how well-suited it may be for LMICs. From here, lower impact and feasibility use cases were deprioritized—for instance, billing improvements use cases received lower priority as they have lower odds of improving patient outcomes and are not as well-matched with low-income contexts.
Once the use cases were prioritized, it became clear that the high-priority use cases seemed to naturally coincide—for instance, the researchers observed a logical link between surveillance and prediction and intervention selection and targeting. This yielded four new use case groupings, which will be discussed in part two of this series: AI-enabled Population Health, Frontline Health Worker Virtual Health Assistants, Patient Virtual Health Assistants, and Physician Clinical Decision Support.
