Potential Supervisor: Matthias Schulte-Althoff, Alessia Nowak

Level: Master

Strategic Orientation: AI & Analytics or Evaluation

Background

Predicting fall events in clinical settings has remained an unresolved problem for years. Validated fall risk screening and assessment tools show only insufficient predictive performance among older hospital patients (Lee et al., 2023; Matarese et al., 2015). Likewise, there is no evidence that FRATs (Fall Risk Assessment Tools) are superior to clinical judgement (Meyer, Möhler & Köpke, 2018; Cameron et al., 2018).

One key reason is the prevention paradox: preventive measures alter the occurrence of falls and therefore also affect the observed performance of diagnostic tools. If a risk scale is applied effectively, the number of actual falls within the high-risk group decreases. As a result, sensitivity and specificity appear to deteriorate (Defloor, 2004; 2005). The validity of commonly used risk assessment methods is therefore systematically underestimated.

Machine learning models have the potential to better capture complex interactions between risk, preventive measures, and clinical factors (Lee et al., 2025). Initial findings from the KIP study highlight this potential. However, the explanatory power of the pilot study is limited by the low fall incidence (2.5% during the period 01–06/2025; 10 falls among 394 patients). An analysis based on a larger set of routine data (n = 4,700 cases from eight high-risk wards) represents one possible solution.

To date, there has been no analysis of how preventive measures (e.g. bundled fall prevention measures) influence model behaviour.

Research Directions

The aim is to systematically analyse the fall prevention paradox in the context of AI-based prediction models. Methods are to be developed and tested that explicitly account for the influence of nursing prevention measures on diagnostic accuracy and model performance.

Example Research Questions:

• How do nursing prevention measures affect the diagnostic accuracy of ML models for fall prediction?

• Does model performance improve when preventive measures are explicitly included as features?

• To what extent do causal models enable the simulation of counterfactual scenarios such as:
  What would the fall risk be without the preventive measure?
  To what extent does the measure reduce the risk?

• What added value is provided by the expanded routine data (preventive measures, fall assessment) from the evaluation study conducted as part of KIP-SDM?

Data Access & Availability:

The dataset from the KIP-SDM evaluation study is already available and includes 394 patients from several wards at Charité Campus Virchow-Klinikum during the period from January to June 2025.
Data points include diagnoses, medical procedures, nursing data, and preventive measures.

It is planned to further enrich the dataset by extracting routine data for all patients from the participating wards from the hospital information system during the observation period.

Project Context:

The study will build on the findings of the completed three-year KIP-SDM research project funded by the BMFTR.

Requirements and Required Skills

To work on this topic, students should have both an affinity for data and an interest in clinical and/or nursing-related questions. In addition, students should possess a basic understanding of statistical methods and fundamental knowledge of data processing or data science, for example using R or Python.

Students should be willing to engage with a complex, real-world problem at the intersection of information systems, nursing care, and AI. Teamwork and the ability to reflect critically are also essential.

References

Cameron, I. D., Dyer, S. M., Panagoda, C. E., Murray, G. R., Hill, K. D., Cumming, R. G., & Kerse, N. (2018). Interventions for preventing falls in older people in care facilities and hospitals. The Cochrane Database of Systematic Reviews, 9(9), CD005465. https://doi.org/10.1002/14651858.CD005465.pub4

Defloor, T., & Grypdonck, M. F. (2004). Validation of pressure ulcer risk assessment scales: a critique. Journal of Advanced Nursing, 48(6), 613–621. https://doi.org/10.1111/j.1365-2648.2004.03250.x

Defloor, T., & Grypdonck, M. F. (2005). Pressure ulcers: validation of two risk assessment scales. Journal of Clinical Nursing, 14(3), 373–382. https://doi.org/10.1111/j.1365-2702.2004.01058.x

Lee, A., Lee, H., & Lee, S. H. (2025). Digital healthcare approaches for fall detection and prediction in older adults: A systematic review of evidence from hospital and long-term care settings. Medicina (Kaunas), 61(11), 1926. https://doi.org/10.3390/medicina61111926

Lee, V., Appiah-Kubi, L., Vogrin, S., Zanker, J., & Mitropoulos, J. (2023). Current cut points of three falls risk assessment tools are inferior to calculated cut points in geriatric evaluation and management units. Muscles, 2(3), 250–270. https://doi.org/10.3390/muscles2030019

Matarese, M., Ivziku, D., Bartolozzi, F., Piredda, M., & De Marinis, M. G. (2015). Systematic review of fall risk screening tools for older patients in acute hospitals. Journal of Advanced Nursing, 71(6), 1198–1209. https://doi.org/10.1111/jan.12542

Meyer, G., Möhler, R., & Köpke, S. (2018). Reducing waste in evaluation studies on fall risk assessment tools for older people. Zeitschrift für Evidenz, Fortbildung und Qualität im Gesundheitswesen, 137–138, 1–7. https://doi.org/10.1016/j.zefq.2018.09.00