Predicting the Uptake of Long-term Care Allowance in Austria Using Pre-benefit Healthcare Data: a Machine Learning Approach

Background Population aging is increasing the demand for long-term care, making it important to better anticipate future entry into long-term care systems and the associated demand for care services. While long-term projections are common, less is known about the extent to which routinely collected administrative healthcare data can be used to predict entry into long-term care systems in the short term. Methods We analyze pseudonymized administrative individual microdata matched to extensive healthcare records from the Austrian public health system for almost 474,000 individuals between 60 and 85 years. Predictive models use recorded healthcare utilization (doctor contacts, hospital stays, diagnoses, drug dispensings, etc.) within a specified "observation period" to classify the first-time receipt of the Austrian Long-Term Care Allowance (LTCA) in a subsequent "event period". We compare different time horizons and levels of detail for the healthcare data to test their effect on prediction accuracy. To deal with the high dimensionality of our data and to prevent overfitting, we estimate models with Lasso, Ridge, and Elastic Net regularization. Model tuning employs fivefold cross-validation on training data, with performance assessed on held-out test data. Results Our findings show that healthcare records provide substantial explanatory power to predict the uptake of LTCA, especially in the short term. For example, in our best performing classification models, the area under the receiver operating characteristic curve reaches values close to 0.9 when evaluated on the test data. Our results also reveal that a finer level of detail in healthcare records does not necessarily lead to better predictions. Apart from old age, the most influential predictors include the frequency of doctor visits and hospital stays as well as diagnoses such as dementia, cerebral infarction, and hypertension. Conclusions Administrative healthcare data can provide useful signals for predicting future LTCA uptake at the population level and may thereby inform projections of demand for long-term care services. Our contribution lies in quantifying the predictive value of such data and identifying the time horizons and data structures that are most informative for anticipating first-time LTCA receipt. Clinical trial registration Not applicable (observational study).