On 28 November, I’ll be presenting my latest research paper for the IFoA on the topic of Using Predictive Analytics to Improve Healthcare Demand Forecasting.
In simple terms, predictive analytics is the process of learning from historical data to make predictions about future unknowns and can involve either supervised or unsupervised learning. The paper looks at how data from a variety of different sources have the potential to be used in predictive analytics models to improve forecasting of demand for health care services. In particular, it looks at examples from 3 broad categories of data and also considers the future uses of these data sets:
- Health system data: data from the health system such as electronic medical records, discharge rates, readmission rates and condition-specific data.
- Everyday data: data from wearables, mobile apps and Internet of Things devices.
- Genomic data: data from large-scale population based genetic sequencing.
If the data sources above can be combined into interoperable multi-source data sets this should allow more powerful predictive analytics models to be built. There may be the opportunity to use the insights from predictive models of health care demand to inform network-level health system strategies. A network-level strategy could be used to help direct patients between community-based or primary care and hospital-based secondary or tertiary inpatient care in line with their medical need and the capacity of the health system.
Predictive models could allow those in the community setting at high risk of requiring medical care in the future to be identified. This could be achieved through predictive analysis of an individual’s personal health record that collates everyday data in real-time from monitoring devices such as wearables and mobile apps. The record would also contain the individual’s electronic medical records and/or genomic data. Based on the dynamic output of a multi-data source predictive model, the individual could be notified of an upcoming health risk. They could then be directed towards primary care before their condition deteriorates or be proactively triaged into secondary care if their condition requires more acute medical care.
In addition, personal health record data could be aggregated up to a higher level to provide insights including patient-feedback of hospital services, community-based sickness trends and overall hospital admissions. Predictive models could be used to identify when demand for secondary and tertiary in-patient care is likely to surge. Resources including staff, medical equipment and pharmaceutical supplies could be increased across the health system to help match demand. If hospitals are predicted to reach capacity, then data from personal health records including from electronic medical records and medical monitoring devices could be used help to identify those patients currently in secondary and tertiary care that are the most stable and have the lowest risk of developing health complications. These patients could then be directed back into primary and community care, with wearable and IoT devices allowing them to continue to be monitored outside of the traditional hospital setting. This would allow the system to free up resources in anticipation of a demand surge for urgent hospital care by prioritising patients in highest clinical need of treatment within specialised medical facilities and discharging patients who could be managed within the community/primary care setting.
Developing a network-level system that responds to health care demand forecasts in this way is likely to require significant investment of time and capital to build a supportive infrastructure. For example, it important when constructing these models to ensure that there is a common data standard in place to allow information to be easily communicated and interpreted. It will also be important to ensure individuals’ data privacy requirements are met. The implementation of functionality such as block chain can help to reduce cyber-security risks.
With a supportive infrastructure in place, predictive models that analyse historical trends and patterns in population-level patient data could be used to produce dynamic forecasts of likely future demand across the health system. If health-care decision makers can successfully interpret this data, then they could build response mechanisms that help to allocate health care resources according to anticipated patient need and that help to prevent health care systems from becoming overwhelmed in times of projected peak demand.
There is a potential role for actuaries to assist in the construction of and interpretation of outputs from such models. Actuaries could also help to manage the risks inherent in the modelling process through quantifying limitations in the modelling process such as data quality and selection bias. Actuaries could work alongside other stakeholders in the modelling process to find ways to overcome these limitations and thereby reduce the uncertainty of a model’s output. In this way, the actuarial community can help to contribute towards the construction and use of insightful models of health care demand. The outputs of models can be used to help direct the use of health services according to an individual’s anticipated clinical need whilst taking into consideration the forecasted capacity constraints of the health care service.