The Future of Healthcare in the Age of AI - Perspectives from our work at CareCentra
By David Kinney, Ph.D Lead Decision Scientist, Vasant Kumar, Founder & CEO at CareCentra
The quintuple aim in healthcare refers to five key goals for improving the US healthcare system. These are:
- Improving the patient’s experience of care
- Improving the health of populations through better outcomes
- Reducing the per capita cost of care
- Improving the work life of healthcare providers, and
- Reducing inequities in care
Achieving these aims requires that clinicians, providers, and patients work together to identify risk as early as possible and prevent it from presenting in drawn out and expensive care processes with uncertain outcomes. My challenge as a decision scientist is to assess the progression of risk at a population level but with the granular detail necessary to manage these risks at an individual level. The next generation of AI and deep learning technologies will allow us to achieve precision at scale by predicting risk progression and identifying upstream risk as never before. This will help healthcare providers to intervene early, yet only when necessary, preventing complications and hospitalizations. In so doing, they will slow the progression of risk and in some cases even turn back the risk curve by shaping patient behaviors in response to risk trajectories – our primary endeavor at CareCentra.
One of the many areas where we deploy deep learning to help detect upstream risk is in women’s health, specifically in preventing pre-term birth (PTB). It has been estimated by the March of Dimes Foundation that 36% of infant deaths in the US
are pre-term related. There has been an increase in PTBs in the US in recent years, costing the system over $25B annually. To put this in context, every day that a woman stays pregnant after 24 weeks saves $10,000 in the cost of taking care of the baby, according to Intermountain Healthcare
Evidence from our work at CareCentra
Our risk-sensing and personalized-nudging app, MyMoBeMapTM, is currently used by a diverse population of expectant mothers with both known and unknown risk of PTB. It collects risk signals and behavior markers across multiple modalities. First, it uses survey responses to measure a patient’s degree of motivation and their ability to manage their health risks, by assigning the patient a ten-component “MoBe Map.” Second, the app dynamically collects real-time data from patient reported outcomes, wearables, sensors, and symptom checkers from remote patient settings. Our deep learning algorithm uses this data, in conjunction with zip-code level social determinants of health, to predict the gestational age of the patient’s child at birth. The algorithm’s predictions are dynamic and self-correcting, beginning at some point in the second trimester. Patients whose predicted gestational age at birth is less than 260 days (37 weeks) are watched closely and nudged constantly with a series of risk-mitigatory actions, including in-person monitoring of the patient’s condition (when triggered by a tiered alerting system), to push delivery as close to 260 days as possible.
This methodology has been rigorously demonstrated to improve patient outcomes. In conjunction with physician collaborators, we participated in a randomized clinical trial
to measure, as a secondary endpoint, the efficacy of MyMoBeMap in preventing the most severe consequences of pre-term birth. Our study found that infants born pre-maturely to mothers who had opted to be on the CareCentra platform tended to spend an average of 6.8 days in the Neonatal ICU, as compared to an average of 45.5 days for infants born prematurely in the control group (p = 0.005).
Technical approach and methods
Implementing a successful patient-centered program like this requires highly flexible technology with sparse data models that can recommend risk-mitigating micro actions to patients on the fly. Our risk-sensing algorithm needs to accept input across a wide range of data modalities, including everything from clinician-generated text to patient-generated wearables data. Moreover, the number of data points varies by patient and our models must be able to accept input of various cardinality. In addition, model output must reflect the inherent uncertainty in medical predictions. To achieve this capability, we combine flexible-input LSTM and liquid neural net layers with a Bayesian dense variational layer to learn a probability distribution over model parameters. This in turn allows us to sample output and generate a reasonable confidence interval providing upper and lower bounds on our predictions.
Why is this important to the future of healthcare?
CareCentra helps to achieve the quintuple aim of care by automating risk detection via a patient-centered application and deep-learning architecture that demonstrates real-world results. Most importantly, the same self-learning Unified Health Algorithm that powers our results in maternity is highly adaptive and addresses several comorbidities from medication adherence to diabetes, from cardiac care to respiratory disease by emphasizing detection of upstream behavioral risk signals in all these cases.