She is 78, lives alone, has mild heart failure and a history of a minor stroke. She has fallen on an outstretched hand and sustained a displaced proximal humerus fracture. In the trauma meeting, the question comes up: operative or non-operative? Her son, who has driven three hours to be there, wants to know whether surgery will help her.
Before you can answer that question, there is a prior one that almost never gets addressed directly: what is her 1-year mortality risk regardless of the management choice?
The study
A paper published in Clinical Orthopaedics and Related Research in January 2026 built a machine learning tool to answer exactly that. The study used a retrospective cohort of 2,999 patients aged 65 and over presenting with first-time proximal humerus fractures to two Dutch hospitals — one Level 1 and one Level 2 trauma centre — between 2016 and 2023. Four ML algorithms were developed: logistic regression, XGBoost, random forest, and LightGBM. They were trained on 1,768 patients from the first centre and externally validated on 1,231 patients from the second, geographically distinct site.
C-statistics ranged from 0.80–0.81 internally and 0.83–0.85 on external validation. For a mortality prediction model in a geriatric trauma population, this is strong — the published literature benchmark sits around 0.75. Negative predictive value was 0.91. Logistic regression was selected as the final model for calibration and interpretability.
What predicts mortality
The three strongest independent predictors of 1-year mortality were hemiplegia, prefracture residence in a healthcare institution, and heart failure. These are data you already have at presentation. The tool consolidates them into a probability estimate using variables that require no additional investigation beyond a standard clerking.
What the tool is for — and what it is not for
Surgical and non-operative treatment were deliberately excluded from the model. The aim was to estimate mortality at the moment of fracture, before any management choice. That is the key design decision, and it determines how the tool should be used.
A 1-year mortality prediction is useful for realistic consent and shared decision-making. It is not a recommendation about whether to operate. The patient with a 35% 1-year mortality risk may still benefit from operative fixation. The patient with a 12% risk may reasonably choose non-operative management. The number informs the conversation — it does not replace the clinical judgment about which intervention best serves that patient’s goals.
The tool is freely available as a web application: bjarty.shinyapps.io/mortality_app
Caveats
Both validation centres were Dutch. Performance held on external validation within that context, but there is no UK data. Differences in comorbidity burden, baseline frailty scoring, and health system context mean calibration may not transfer identically. The tool applies to patients confirmed to have a first-time proximal humerus fracture — it is not a screening tool for undifferentiated falls presentations.
The practical implication
Before the management discussion in any elderly proximal humerus fracture patient, use this tool to put a number on 1-year mortality risk. Not as a substitute for clinical judgment — as a structured input to the consent conversation. “Based on your medical history, the model estimates your risk of dying within a year of this injury at X%” is a different and more honest discussion than the implicit assumption that fracture management will be the main determinant of survival.
In an 80-year-old with heart failure and a prior stroke, it often is not. Making that explicit, at the start of the conversation rather than after the operation, is what this tool is for.
References
- Mennes SR, et al. Machine learning-driven probability calculators can accurately predict 1-year mortality after proximal humerus fractures in patients over the age of 65 years. Clin Orthop Relat Res. 2026. PMID 41564299. https://doi.org/10.1097/CORR.0000000000003828