+66% growth in ART & PrEP users since 2016.
Growing pressure on clinical care requires more efficient use of resources.
AI can help to optimize clinical care and transition clients to non-clinical care.
The global effort in ending the HIV epidemic has come a long way. With breakthroughs in medical science over the past 20+ years we have seen HIV/AIDS move away from being a fatal illness to a lifelong chronic condition. What had once been a race against time has now become a challenge of access and distribution.
With more people living longer, fuller lives while living with HIV, the need to adapt services and to transition clients rapidly and safely to care and treatment that works for them - such as non-clinical care - is becoming more urgent.
In this blog post we discuss the challenges of the next stage of the HIV epidemic and present evidence on how AI can not only improve current clinical distribution systems, but also prepare for non-clinical differentiated care and service delivery models (DSD).
Over the last two decades HIV treatment and prevention medication, namely antiretroviral therapy (ART) and pre-exposure prophylaxis (PrEP), have seen significant improvements. To maximize effectiveness of medications, clients are initiated earlier on ART or PrEP. These regimens require anywhere between one to twelve clinical appointments and two to twelve pharmaceutical pick-ups a year for the rest of their lives.
Retention in care has shown to be crucial to not only minimize deaths and new infections, but also to maximize resources. Those who do drop out of treatment risk a resurgence in viral load, and if not successfully reinitiated in care may have developed drug resistance, requiring more aggressive medication.
Considering these outcomes, in order to achieve the second and third of UNAID’s 95-95-95 targets, retention in care is crucial.
While innovations in treatment, prevention, and tracking have significantly advanced our efforts to combat the epidemic, they are now presenting new challenges due to their own success.
As we initiate more people on treatment regimens and life expectancies go up, so are the number of people that continuously require care and follow-up to avoid interruption in treatment (IIT). Hospitals and health clinics are being stretched beyond their capacities as new clients are initiated while existing clients remain in care longer.
Furthermore, with HIV no longer being considered a public health threat in many high-income countries, resources to fight the epidemic in low- and middle-income countries are shrinking. Funding has dropped to 2013 levels and UNAIDS is estimating an annual funding gap of $9bn. This comes in addition to an estimated 15% of people who still need to be initiated on ART, and an estimated 77% to PrEP, by 2025.
With care providers stretched, funding shrinking, and ART and PrEP clients expected to be initiated on treatment for longer periods of time, two important questions emerge regarding the path forward. How do we ensure that we meet and sustain the second and third 95 targets? And, how do we prepare for the initiation and delivery of care and treatment services to an additional 13.4m people using less resources?
This comes down to minimizing IIT in already initiated clients, and evolving distribution systems and protocols into scalable, decentralized non-clinical care.
Retention in care is not linear. Throughout a client’s journey on ART, individuals cycle in and out of care based on environmental and behavioral factors. Current interventions designed to prevent IIT are reactive and costly. To achieve epidemic control, interventions need to be predictive and proactive. This is where AI and machine learning have proven to be valuable tools.
In recent years we have seen promising research on a number of successful pilots and studies in HIV programs using AI as a tool for health care workers to better predict and prioritize clients at risk of IIT. In a study conducted in South Africa in 2022, researchers used machine learning to correctly classify risk of non-adherence in two out of three patients and viral suppression in three out of four patients. In another study, conducted by ICAP in Mozambique in 2018, machine learning was successfully used to identify a subset of patients at higher risk of default. Now that we know the technology works, we need to deploy it at scale to improve clinical outcomes in initiated clients.
Pendulum is a technology company that applies AI onto supply chains to improve demand forecasting and supply allocation. Its product Predict\People has been deployed in clinical settings to improve client outcomes since 2017. It leverages existing health system data to learn a client's likelihood of experiencing IIT with high precision. The first version of the product was deployed in a USAID-funded project in partnership with Data.FI. Predict\People showed that it could predict IIT with unprecedented accuracy at the clinical or drug pickup appointment-level. The results were published in the Journal of Acquired Immune Deficiency Syndromes in 2022.
However, similar to any other AI solution deployed with real users, the prediction models do not encompass the whole solution. In order for AI to be useful beyond research, it needs to be actionable for the intended users. This requires the engineering and orchestration of several supporting pieces. For example, Pendulum\Predict is built using infrastructure that ensures both functionality and scalability. The product’s backend allows models to be run in real-time on large quantities of existing patient data. This prevents long waiting times and friction in the user flow. Moreover, Pendulum integrates Predict\People’s predictions into clinical care workers’ existing software platforms and workflows, eliminating the need for training or user handling errors.
With Predict\People, health workers continue with the same process they are used to, but become more effective with the help of AI. Client lists are generated as before in under 30 seconds, only now the names appear in order of highest risk of IIT. This allows the prioritization of clients with the highest risk for additional intervention before a missed appointment. Health workers are enabled to achieve maximum possible impact, even as appointment lists become longer.
Lastly, AI outputs need to be accessible and interpretable by the end user. Pendulum deploys Predict\People on end user devices with limited connectivity via patient or appointment management systems, for example, an OpenMRS module. This makes outputs available not only for central decision making on resource allocation, but also in real-time at the operating edge where health care workers make decisions on when and how to intervene.
The current approach to end the epidemic scales linearly with the problem. The faster we want to end the epidemic, the more people we need to onboard and keep on ART and PrEP, the more clinical care we need and the more resources we need. Building and maintaining physical infrastructure to solve the epidemic is not only costly, but also brings about the question on what to do with it once we have won the fight. Finding ways to do the same with bytes and software would be faster and more resource efficient.
Today clinical care is the default choice as a distribution system because IIT is at the core of the problem. Health organizations and clinical workers want to see the clients in person and make sure they are taking their medication, and rightfully so. Until now we did not have the tools to understand who is at risk and should be followed up with through clinical care versus who is not and can be trusted with non-clinical care.
Products like Predict\People change that. While in their current form they output client lists sorted from high to low IIT-risk for follow-up, the prioritization is powered by risk scores. Using these IIT risk scores hospitals and HIV clinics can gain granular real-time visibility into at-risk populations at a facility-level. This allows them to segment the population and provide DSD.
DSD can reduce the pressure on clinical care and direct resources to onboard and maintain more ART and PrEP clients with slowing funding, instead of building rigid physical infrastructure that is hard to scale up and down.
Differentiated non-clinical care becomes even more valuable when we consider the positive impact that meeting clients’ needs better can have on adherence and retention. With clients living longer and healthier lives, the inconvenience and stigmatization associated with regular clinical visits or pharmaceutical pickups continue to accrue. When considering ‘appropriate’ care and treatment over the long term, normalcy and the ability to live full lives should become part of the calculation.
The current stage of the global HIV response relies upon maximizing client adherence and utilizing existing resources optimally. Predict\People is a proven tool being used by health systems to reach the 95-95-95 targets, improving patient adherence and health despite growing case loads and declining resources.
Reach out to Pendulum today to learn more about how Predict\People can segment patients optimally for care and treatment options and reduce your health workers’ workloads by focusing them on the interventions that matter most.
Pendulum is an AI company that optimizes critical supply and demand networks. We make ubiquitous systems more intelligent, maximizing the impact of resources available. Our products are deployed via APIs that continuously predict demand, optimize supply, and improve on their own.
If you’d like to learn more about Predict\People or Pendulum, please send an email to [email protected].