Predicting healthcare needs and outcomes could radically transform the future of the sector, argues Lisa Balboa
An increasing amount of data is becoming available for actuaries to use in forecasting health. It is already being leveraged to predict hospital admissions, improve health outcomes and forecast epidemics. Actuaries can scale up these uses by working to improve data quality and interoperability and overcome privacy challenges. Healthcare-specific algorithms can then be used to forecast morbidity across entire health systems.
Three broad categories of data can be used in health forecasts. The first is health system data: traditional health service data sources, such as electronic medical records, discharge rates, admission rates and condition-specific data. The second is everyday data: data from wearables, mobile apps, social media and Internet of Things (IoT) devices. The third is genomic data: large-scale population-based genetic sequencing, including direct-to-consumer genetic testing kits and national genomic sequencing programmes. There are already localised examples of how these data sources are transforming healthcare.
Predicting hospital admissions rates
Hospital admissions data is being used to forecast demand for inpatient admissions. The Patient Admission Prediction Tool has been developed to forecast hourly patient arrivals by categories such as gender and medical condition across 27 hospitals in Queensland, Australia. The model uses more than 10 years of patient admissions data, along with other attributes, such as times of year when infectious disease outbreaks are most likely. This has enabled hospital bed managers to better plan resources for the coming week by considering the level of predicted inpatient arrivals alongside current hospital occupancy levels.
Big data is also being used to help forecast epidemic outbreaks. International researchers have developed a model that can predict cholera outbreaks in Yemen up to four weeks in advance; data inputs include rainfall radar forecasts, population density, access to clean water and seasonal temperature. Health workers are sent to areas with the highest likelihood of an imminent cholera outbreak, and help to reduce the spread of disease by providing sanitation advice, chlorine tablets and a range of other health measures. If the model's prediction window could be doubled, vaccination campaigns could also be proactively targeted. Going forward, researchers want to extend the model to more regions in order to predict a wider range of diseases, including dengue fever and malaria.
Improving health outcomes
Genomic data is also emerging as a big data source for health prediction, and institutions are striving to sequence millions more genomes. Launched in late 2018, the NHS Genomic Medicine Service is aiming to sequence one million genomes by encouraging individuals to share their genomic data for the benefit of the research community. The long-term goal is to use deep learning and other techniques to establish the genetic basis for diseases such as cancer. Using these insights, personalised medical treatments could then be targeted based on a patient's genetic profile.
Big data challenges
These examples show the benefits that are emerging from the use of big data in predictive models. More value could be derived if models incorporate a wider range of diseases and health forecasting is rolled out nationally. However, challenges must be overcome before models can be scaled up to achieve this.
The predictive power of models could be improved by creating interoperable multi-source datasets. International projects such as Fast Healthcare Interoperability Resources (FHIR) are aiming to define a common data standard for the exchange of electronic healthcare information. In the UK, the NHS is striving to improve the sharing of digital health records by adopting the FHIR standard. In the longer-term, health service, everyday and genomic data could also be standardised and pooled together. This would allow a more representative view of individuals' health to be captured both in the clinic and as people go about their daily lives.
Another challenge is data privacy. Healthcare models need to be compliant with data protection legislation, and public trust must be earned to ensure individuals are willing to share their sensitive health and lifestyle data. A potential solution to this could be the implementation of personal patient health records. Individuals would be the gatekeepers of their data, choosing exactly which third parties could access which parts of their health system information, as well as their everyday and genomic data. An app-based solution that uses blockchain-type technology could help to secure data access. Individuals could also authorise the de-identification of their data for use in national research programmes and health forecasts.
Actuaries also have a leading role to play in the development of healthcare-specific algorithms. Off-the-shelf machine learning algorithms often aren't designed to analyse health data, but operations in the healthcare industry; for example, certain invasive medical tests are used primarily to confirm the doctor's suspected diagnosis, rather than being predictive of the disease itself. To overcome systemic biases in the data, the development of large-scale predictive healthcare models will require thoughtful design and collaboration with medical professionals.
Predict and survive
If these challenges can be overcome, insights from predictive models of healthcare demand could be used to inform network-level health system resourcing decisions. Individuals could be directed to primary and hospital care according to the likely severity of their condition. They would also be sent to facilities that have availability, in order to avoid lengthy waiting times for treatment.
For those not currently receiving any care, data from their personal health record could be analysed. Preventative lifestyle and health measures could be suggested to reduce their likelihood of needing medical care in the future. Additionally, the model may proactively suggest healthcare is needed by detecting signs of disease in their wearables and IoT data.
When an individual needs care, they can be triaged using insights from their personal health record. If the condition is predicted to be severe, they will be directed straight to hospital care. If it's predicted to be less acute, primary care will be suggested. Moreover, by aggregating individuals' data at a higher level, demand surge in the system could be forecast and bottlenecks prevented. For example, if a severe winter flu outbreak is predicted, resources could be proactively increased across the health system to help meet demand.
If the data forecasts that a hospital is due to reach capacity, even once resourcing has been scaled up, the most clinically stable patients who are at lowest risk of developing health complications could be directed back into primary care. Wearables and other telehealth devices could be used to monitor their condition under the care of doctors and other medical professionals in the community. This would free up inpatient hospital resources, making them available for the critically ill, who require more specialised medical care.
Opportunities for actuaries
Developing a network-level strategy that leverages big data to forecast healthcare needs will require significant investment. Actuaries should help to develop a supportive infrastructure that overcomes data interoperability and privacy risks. Health system, everyday and genomic big data sources can then be aggregated and used as dynamic inputs in responsive models of healthcare demand.
Actuaries can communicate the outputs of these models to key stakeholders, facilitating the optimisation of healthcare resources. Preventative care can be targeted more effectively to reduce the chronic disease burden and proactively limit the spread of infectious diseases. Individuals can be more efficiently directed to medical care by considering their anticipated clinical need alongside the forecasted capacity constraints of the health service.
Lisa Balboa works at Bupa and is on the managing committee of the IFoA's Data Science Member Interest Group. The opinions expressed in this article are the author's own views.