The Power of Heuristics in Predicting Fracture Nonunion

Abstract

Background/Objectives: Although extensive research on risk factors for nonunion development has been published, clinicians frequently rely on heuristic reasoning-intuitive, experience-based decision-making-to predict nonunions. However, the accuracy of these intuitive assessments and the influence of clinician experience remain uncertain. This study aims to assess clinicians' diagnostic accuracy in predicting nonunion, investigate the impact of experience on predictive performance, and identify patient-specific factors contributing to diagnostic errors. Methods: This retrospective, multi-center cohort study included 98 patients with surgically treated tibial shaft fractures between 2018 and 2023 from four level-one trauma centers in Germany. Fracture outcomes were classified as either nonunion (n = 20) or regular fracture healing (n = 78). Patient cases were presented to 24 clinicians. Each clinician independently assessed preoperative and postoperative biplanar X-rays and patient histories to predict fracture healing. Results: Clinicians' sensitivity significantly improved from 50.4% to 60.2%, while specificity declined (74.0% to 70.7%) with the addition of postoperative information. No significant differences in predictive performance were observed across different levels of clinician experience. Changes in assessment after reviewing postoperative information were equally likely to be beneficial or detrimental. Certain patient factors, including obesity and smoking, influenced prediction errors. Conclusions: This study is the first to assess heuristic reasoning in nonunion prediction. The findings suggest that clinician experience does not significantly enhance diagnostic accuracy under limited-information conditions. Patients should be informed that predicting individual nonunion risk remains challenging. Larger studies are needed to explore the role of patient-specific factors and refine clinical decision-making in fracture healing prognosis.

Keywords: cognitive biases; fracture healing; heuristic reasoning; nonunion.

Figure 1

Demonstration of a patient case as presented to the clinical raters for healing prediction: (a) preoperative biplanar X-rays and patient background information; (b) postoperative biplanar X-rays (original text in German). R = right side.

Figure 2

Fracture types according to the AO classification.

Figure 3

Performance ratings of all raters combined: Postoperative information like surgical reduction and implant information significantly increases the sensitivity and reduces clinician raters’ specificity. Sensitivity_preop = 50.4% vs. Sensitivity_postop = 60.2%, p < 0.001; Specificity_preop = 74.0% vs. Specificity_postop = 70.7%, p = 0.08; PPV_preop = 38.1% vs. PPV_postop = 35.5%, p = 0.40; NPV_preop = 85.8% vs. NPV_postop = 87.7%, p < 0.001. *** = p < 0.001. Error bars: bootstrapped 95% confidence intervals.

Figure 4

Performance ratings of raters categorized by professional position: (a) Preoperative evaluation. No significant differences were observed between positions. (b) Postoperative evaluation. No significant differences were observed between positions. (c) Delta between preoperative and postoperative evaluation. Postoperative information, such as surgical reduction and implant details, led to a significant increase in sensitivity, particularly among specialists and residents, but also reduced their specificity. Additionally, the NPV of specialists showed a significant but marginal increase. However, no significant differences were observed in the delta between preoperative and postoperative evaluations across the categorized positions. ++ = p < 0.01, + = p < 0.05, preoperative to postoperative metric. Error bars: bootstrapped 95% confidence intervals.

Figure 5

Correlation between raters’ performance metrics and experience. No significant correlation was found between performance metrics and rater experience. Preoperative PPV showed the highest correlation with experience but remained insignificant (Pearson’s r = 0.386, p = 0.062): (a) preoperative evaluation; (b) postoperative evaluation. Shaded areas indicate 95% confidence intervals.

Figure 6

Violin plots illustrating changes from preoperative to postoperative ratings after receiving additional information. The plots display the total number of changes and the counts of beneficial and detrimental changes for each individual clinician (black dots).

Figure 7

Mean prediction error by clinicians, stratified by the presence (1) or absence (0) of clinically relevant patient factors. Positive prediction errors reflect a tendency to overestimate the factor’s impact (false positives), whereas negative prediction errors indicate underestimation (false negatives): (a) compartment syndrome; (b) open fracture; (c) polytrauma; (d) post-surgical fracture gap (>4 mm); (e) obesity; (f) smoking. Error bars: bootstrapped 95% confidence intervals. Note the adjusted y-axis scales in (e,f) to better fit the data.