Immediate-postoperative risk stratification after PFNA centers on hidden blood loss
Hidden blood loss after proximal femoral nail antirotation, or PFNA, remains a practical hazard in older patients treated for intertrochanteric fracture. The bleeding is often not obvious in the operating room, yet it can drive postoperative anemia, delay recovery, and complicate transfusion decisions when patients are already physiologically fragile. This retrospective prediction study addressed that problem with a narrowly focused question: whether an immediate-postoperative model could identify patients at high risk for hidden blood loss using information available before and during surgery.
The investigators built the model in 186 older patients and tested it in an independent external cohort of 84 patients. High hidden blood loss was defined as more than 650 mL, using an X-tile-derived cutoff linked to postoperative anemia as the clinical endpoint. In the development cohort, 99 of 186 patients met that threshold, a rate of 53.2%. In the validation cohort, 27 of 84 patients did so, or 32.1%.
A single threshold is clinically useful only if the model can discriminate risk without exaggerating it. In this study, that balance was uneven but informative.
Three predictors carried the final model
The final immediate-postoperative model retained three variables: preoperative anticoagulant use, preoperative albumin, and operation time. Each has a plausible physiologic link to perioperative bleeding burden. Anticoagulant exposure can amplify blood loss and make a modest operative bleed translate into a larger hidden deficit. Low albumin often tracks frailty, malnutrition, inflammation, or reduced physiologic reserve, all of which may correlate with worse surgical tolerance and poorer hemostatic compensation. Operation time is a direct marker of technical complexity and tissue trauma; as the procedure extends, cumulative oozing from bone and soft tissue can continue well after the visible surgical field seems controlled.
The apparent development area under the curve was 0.715, with a 95% confidence interval of 0.640 to 0.790. Internal validation used 1,000 bootstrap resamples, and the full modeling workflow was repeated within each sample, including LASSO tuning. After optimism correction, discrimination fell to an AUC of 0.660, the optimism-corrected Brier score was 0.239, and the optimism-corrected calibration slope was 0.735. Those numbers point to some overfitting inside the derivation cohort, even before the model is applied elsewhere.
External validation delivered stronger discrimination but imperfect calibration
The independent external cohort told a different statistical story. Discrimination improved sharply, with an AUC of 0.894 and a 95% confidence interval of 0.826 to 0.963. The Brier score was 0.165, and calibration remained close enough in slope terms, at 1.026, to suggest the ordering of risk held up well across centers. Yet the calibration intercept was -1.806, a sign that the model overestimated absolute risk in the external setting.
The distinction matters at the bedside. A model can rank patients well and still assign probabilities that are too high for the population in front of the surgeon.
After intercept-only recalibration, the Brier score improved to 0.121 while AUC remained unchanged. In practical terms, the model’s risk ordering survived transfer to a new cohort, but its raw probabilities needed local adjustment before they could guide transfusion planning, monitoring intensity, or postoperative anemia surveillance. This is a common failure mode for clinical prediction models built from modest sample sizes: transportability may be better than calibration, and clinicians can overtrust the numeric risk estimate if they do not recalibrate it to their own patient mix.
A parsimonious model can help if it stays local
The authors also reported a supplementary preoperative model, which achieved an external AUC of 0.839. That comparison is notable because it suggests that risk information exists even before the operation begins, but the immediate-postoperative version was designed for a different workflow. It was meant to support decisions right after surgery, when a team is deciding how closely to follow hemoglobin trends, whether to anticipate transfusion, and how aggressively to prepare for anemia-related instability.
This is where hidden blood loss becomes more than a statistical endpoint. After PFNA, the visible operative blood loss can underestimate the true physiologic burden by a wide margin because blood can sequester into tissues, fracture sites, and the medullary compartment. Older patients tolerate that mismatch poorly. Reduced reserve, lower albumin, and anticoagulant exposure can all magnify the impact of a blood loss that would be less consequential in a younger, healthier person.
The study’s design also deserves attention. It used retrospective de-identified data collected from 2018 to 2023, with ethics approval granted for the analytic study in 2026. The work was supported by a single institutional research project, and no competing interests were declared. Those details do not change the model’s performance, but they do frame the evidence base as a center-specific effort rather than a multicountry validation program.
Why the model should not be used as-is everywhere
The cleanest lesson from the validation data is also the most restrictive one: the model is best treated as a local risk-stratification tool, not a universal clinical rule. Strong external discrimination does not erase the fact that the model substantially overestimated absolute risk in the validation cohort. Without recalibration, a clinician could overcall danger, intensify monitoring unnecessarily, or make transfusion decisions on inflated probabilities rather than on actual physiology.
Future use depends on two steps already named in the study itself. First, local recalibration is needed wherever the model is deployed. Second, prospective multicenter validation is still required before broader implementation can be justified. Those are not formalities; they are the conditions that determine whether a prediction tool becomes useful at the bedside or remains a promising derivation exercise.
For older patients undergoing PFNA for intertrochanteric fracture, hidden blood loss is common, clinically meaningful, and easy to miss. This model offers a structured way to anticipate it immediately after surgery, but the external calibration gap means it should inform judgment rather than replace it.