Reducing Hospitalisations Through Risk Prediction 

• In ICU
• Mon, 7 Jul 2025

Preventing unplanned hospital admissions among older adults is an ongoing challenge for hospitals seeking to maintain patient safety while optimising resource use. Emergency admissions are associated with adverse outcomes, including cognitive decline, hospital-acquired infections and repeat admissions. Understanding the factors that contribute to such admissions is critical for care teams and hospital managers. 

Recent research has demonstrated the value of machine learning (ML) models in identifying these risk factors. A new study based on UK Biobank data has validated several predictors of 30-day emergency hospitalisation, particularly inappropriate polypharmacy as measured by the Drug Burden Index (DBI). The findings suggest opportunities for hospital systems to incorporate validated risk factors into clinical decision-making, supporting early intervention and improved care planning for older patients. 

Validated Risk Factors and Their Clinical Relevance 
The study focused on older adults aged 65 and above, drawing on a cohort of 86,870 participants from the UK Biobank. Participants who had been hospitalised in the year prior to the index date were excluded to minimise healthy user bias. Among the remaining population, 715 individuals experienced an emergency hospital admission within 30 days of the index date. Predictors included demographic details, comorbidities, geriatric syndromes and the presence of DBI medications. The DBI identifies exposure to drugs with anticholinergic and sedative effects and serves as a measure of potentially inappropriate polypharmacy. 

The validation of DBI as a key predictor across all three ML models—Random Forest (RF), XGBoost (XGB) and Logistic Regression (LR)—reinforces its relevance in hospital settings. Other modifiable factors validated by the models included hazardous alcohol use, smoking, mobility issues, falls and fractures. Geriatric syndromes such as dizziness and urinary frequency were also significant. The inclusion of deprivation measures, such as the Townsend index, further highlights the importance of social determinants of health in predicting hospital risk. Although interventions to reduce DBI exposure have had varied success, they remain a promising strategy for decreasing emergency hospital admissions in older adults. 

Implications for Decision Support and Clinical Workflows 
The use of machine learning models in this context offers practical benefits for hospital-based decision support. The study employed structured data readily available in electronic health records or obtainable through direct patient assessment. The XGB model achieved the highest predictive accuracy, with an area under the receiver operating characteristic curve (AUC-ROC) of 0.863, followed by the RF model with 0.785. These models demonstrated good discrimination in identifying individuals at high risk of short-term emergency admission. While the LR model showed lower predictive performance, it still validated key predictors consistent with the other models. 

The models’ reliance on simple binary variables enhances their feasibility for clinical deployment. Features such as presence or absence of DBI medications, history of falls or fractures and mobility limitations could be used to inform discharge planning or prioritise patients for geriatric assessment. Hospital pharmacists and clinicians could apply these insights to conduct targeted medication reviews or initiate preventative interventions. Additionally, risk stratification tools based on these models may support operational planning by identifying patients at increased risk of readmission, thereby improving resource allocation and bed management. 

Must Read: Reducing Hospital Length of Stay Through Innovation 

Despite these strengths, the models also revealed limitations that must be addressed before clinical integration. Calibration issues at the extremes of risk distribution affected the reliability of predictions for low and high-risk patients. These issues must be resolved before the models can be safely deployed in real-time clinical workflows. Moreover, the study cohort included a narrow age range of 65 to 70 years and featured healthier volunteer participants, which may limit generalisability to more clinically complex populations. The use of a fixed index date also constrained the model’s flexibility and future work may benefit from a rolling index approach. 

Implementation Challenges and Future Directions 
Although the predictive performance of the XGB and RF models was strong, their integration into clinical systems will require further refinement. Missing or low-quality data could reduce predictive reliability, and the models will need to be tested on unbalanced, real-world datasets to ensure practical utility. However, the foundation laid by this study is significant. The findings support the inclusion of validated risk factors—particularly DBI, alcohol use and mobility issues—in hospital screening tools and predictive analytics platforms. 

Hospitals should consider supporting the development of decision support systems that incorporate these findings. By identifying patients at risk of short-term emergency hospitalisation, hospitals can take steps to prevent readmissions, reduce patient harm and improve care transitions. The current study followed the TRIPOD-AI checklist for transparent reporting and made all code and data openly accessible, promoting reproducibility and adherence to FAIR principles in artificial intelligence research. 

Hospital systems aiming to reduce emergency admissions in older adults can draw on robust evidence linking modifiable risk factors to short-term hospitalisation. The validation of DBI and other predictors in a large UK cohort supports their relevance in routine hospital care. Integrating such variables into clinical decision-making may allow for more targeted interventions and improved patient outcomes. While model deployment requires further optimisation and broader validation, this study offers a solid foundation for embedding predictive tools into hospital care pathways. Improved identification of at-risk individuals could contribute to reduced admissions, better allocation of resources and enhanced quality of care for older adults. 

Source: Age and Ageing 

Image Credit: iStock

References:

Olender RT, Roy S, Nishtala PS et al. (2025) Potentially inappropriate polypharmacy is an important predictor of 30-day emergency hospitalisation in older adults: a machine learning feature validation study. Age and Ageing, 54(6): afaf156