Can five preoperative signals predict kidney injury after hip-fracture surgery?

The answer: moderate internal prediction, not a deployable bedside tool

Five routinely available preoperative signals helped a machine-learning model separate older hip-fracture patients who did and did not develop acute kidney injury after surgery in a retrospective Chinese hospital cohort. Across repeated internal splits, the model's mean area under the receiver-operating-characteristic curve was 0.7231. That is a potentially useful research signal, but it is not evidence that the model is ready to direct monitoring, fluids, medication changes or surgical timing.

The study had only 54 acute-kidney-injury events, selected the best-performing algorithm using part of the same single-centre dataset and did not test the final workflow at another hospital. Its research web application should therefore be read as an implementation of the study, not as a clinically validated calculator. A clinician would still need ordinary assessment, and a hospital would need external validation and a prospective pathway before considering use.[1]

Who was studied and how kidney injury was defined

The cohort covered 616 patients aged at least 60 who underwent surgery for an intertrochanteric femoral fracture at Fujian Medical University Affiliated Zhangzhou Hospital between January 2015 and December 2025. The median age was 83, 388 patients, or 63.0%, were women, and 54 patients, or 8.8%, developed the primary outcome. All had a preoperative creatinine value and at least one postoperative value within seven days; 542, or 88.0%, were checked within 48 hours.

The researchers used the serum-creatinine part of the 2012 KDIGO definition: an increase of at least 0.3 milligrams per decilitre within 48 hours or an increase to at least 1.5 times baseline within seven days. They did not apply the urine-output criterion because consistently timed records were unavailable. Baseline was the lowest creatinine value between admission and surgery, a choice that may reduce transient dehydration effects but may also reflect early fluids or treatment rather than premorbid kidney function.

That definition matters. When the team instead used the first creatinine measurement after admission, 43 rather than 54 patients were classified with kidney injury. Eleven of the 54 primary cases changed category, and the held-out AUC fell from 0.795 to 0.706. The apparent performance was therefore sensitive to how the outcome's starting point was chosen.[1]

What the model used and what it was compared with

The investigators began with 41 demographic, perioperative, comorbidity and laboratory variables. They allocated 432 patients, including 38 events, to development and 184 patients, including 16 events, to a held-out algorithm-comparison group. Imputation, preprocessing, feature selection and tuning used development data. Seven algorithms were compared, including logistic and penalised regression, random forest, support-vector machine, XGBoost and Naive Bayes.

The selected Naive Bayes model used baseline serum creatinine, blood urea nitrogen, chloride, uric acid and a recorded history of chronic kidney disease or renal insufficiency. These are indicators of renal reserve and biochemical state, not novel causal discoveries. The authors used SHAP values to describe fitted predictions but warned that attribution does not establish an independent biological effect or a treatment target.

Missingness was material for some candidates: operative duration was missing for 58.6% of patients, body-mass index for 37.3%, D-dimer for 28.1% and uric acid for 7.8%. Multiple imputation assumed that missingness could be explained by observed information. The final model did not retain the first three highly missing variables, but the assumption still needs testing in new data.[1]

The performance numbers need their denominators

In the original 184-person comparison group, Naive Bayes had the highest observed AUC at 0.795, with a 95% confidence interval from 0.698 to 0.892. Because that same comparison selected the winning algorithm, this number is susceptible to selection optimism. Across 100 new stratified 70/30 partitions, mean validation AUC was lower at 0.7231, with a standard deviation of 0.0260. A complete-pipeline bootstrap that repeated feature and algorithm selection produced an optimism-corrected AUC of 0.756.

At the selected calibrated threshold, 89 of 184 patients were classified as higher risk. Fourteen of 16 cases were detected, but 75 people without the outcome were also flagged. Sensitivity was 87.5%, specificity 55.4%, positive predictive value 15.7% and negative predictive value 97.9%. Most alerts at that threshold were false positives in this cohort.

The complete-pipeline corrected Brier score was 0.084, with calibration intercept 0.02 and slope 1.70. A slope that far from one signals that predicted risks still need attention. Decision-curve analysis suggested possible net benefit across a range of thresholds, but a retrospective curve cannot show that clinicians will respond appropriately or that patients will benefit.[1]

What this could mean for patients and hospital teams

A carefully validated tool might help a perioperative team identify which older fracture patients merit closer review of hydration, haemodynamics, kidney function and potentially nephrotoxic medicines. The inputs are already available in many hospitals, so the burden may be lower than for a model requiring new imaging or biomarkers. It could support a conversation about risk; it should not delay urgent fracture care or make treatment decisions by itself.

The false-positive count shows the operational trade-off. Flagging roughly half the comparison cohort to catch most of 16 events could increase blood tests, monitoring and clinical reviews for many people who would not develop the outcome. Whether that burden is justified depends on which action follows an alert, its costs and harms, and whether earlier intervention changes outcomes. None of those questions was tested here.

For nurses, surgeons, anaesthetists and geriatricians, explainability should mean more than showing a SHAP chart. Teams need responsibility for reviewing alerts, data-quality checks, thresholds suited to local capacity and a way to override the model. Patients and families need to know that a score expresses uncertainty and is not a diagnosis.[1]

Evidence limits, funding and disclosures

This was a retrospective study at one hospital across eleven years. Practice, laboratory timing and records can change over that period. Only 54 events supported development and evaluation, no outside cohort was used, and the best algorithm was selected using the nominally held-out comparison set. Repeated partitions and bootstrap correction are useful safeguards, but remain internal validation.

Outcome ascertainment omitted urine output and depended on routine creatinine testing. Medication histories, including potentially nephrotoxic exposure, and historical ASA classifications could not be reconstructed consistently after information-system changes. Sensitivity to the baseline-creatinine definition shows that a technical choice can change both the event denominator and performance estimate.

The authors reported no financial support and no commercial or financial conflicts. The ethics committee approved the retrospective analysis and waived individual consent. Patient-level data are available only on reasonable request. The paper disclosed using OpenAI Codex for translation, language editing, organisational revision, code review and figure assembly; the authors said it did not generate patient data or independently determine findings and that they verified outputs.[1]

What would change the assessment

Confidence would rise if the frozen pipeline were validated prospectively at multiple hospitals, with prespecified thresholds and complete reporting of calibration, false alerts and subgroup performance. A strong study would compare it with a simple conventional score and clinician judgment, not merely with other algorithms, and preserve a genuinely untouched external test set.

The decisive evidence would be a prospective impact study: does score-guided care reduce clinically meaningful kidney injury, dialysis, complications or length of stay without delaying surgery or creating excessive monitoring? The assessment would weaken if calibration fails elsewhere, if a simpler model performs as well, or if extra surveillance does not improve outcomes.[1]