Recalibrating the Hester Davis Scale for Fall Risk in Patients with Breast Cancer

Introduction

A high number of patients with cancer, especially older adults, experience falls during or after chemotherapy (1–3). In older individuals, the risk of falling is already elevated due to age related declines in balance, strength, and sensory function (3,4). Cancer treatment adds fatigue, neuropathy, and other toxicities (5,6). For women, especially those with cancer, the intersection of cancer, hormone changes, and menopause leads to an even greater fall burden, nearly double that of their male counterparts (7). The Hester Davis Scale (HDS) is widely used in United States hospitals and was developed for a neuroscience hospital unit (8,9) and it has not been calibrated for oncology. To address this gap, we evaluated the HDS in its original form without modification, to determine whether it predicts falls in women with breast cancer and whether the published thresholds require recalibration.

Methods

We evaluated a retrospective cohort of 921 patients with breast cancer who were 50 years or older and received care at a comprehensive cancer center from 2022 to 2024. This study was reviewed by the Virginia Commonwealth University Institutional Review Board (Protocol HM300000100) and was determined to be non-human subjects research. We created a structured dataset that included the nine HDS domains (age, fall history, mobility, medications, mental status, toileting needs, volume and electrolyte status, communication and sensory status, and behavior) obtained from clinical records, and we scored each domain using the original rubric.

The primary outcome was one or more documented falls within one year after diagnosis. The principal predictor was the HDS total score, analyzed as a continuous measure as well as a categorical variable. We summarized cohort characteristics, compared scores between patients who fell and those who did not, and compared fall rates across risk groups. We assessed operating characteristics at the published cut score of 10 and across cut scores from 0 to 11, reporting sensitivity, specificity, positive predictive value, negative predictive value, and the area under the receiver operating characteristic curve. We selected the optimal threshold using the Youden index (10). We also used logistic regression with at least one fall as the outcome and age, race, ethnicity, cancer stage, and treatment exposures as covariates. Groups were compared using two-sided t-tests with alpha equal to 0.05.

Results

Demographic characteristics.

The cohort included 921 patients with breast cancer. The mean age at diagnosis was 66.2 years with a range from 50 to 89 years. The first quartile was 60 years, the median was 66 years, and the third quartile was 72 years. At data capture, 892 patients were alive (96.9%) and 29 were deceased (3.1%). By race, 583 patients were White (63.3%), 302 were Black (32.8%), seven were Other (0.8%), and 29 were Unknown (3.2%). By ethnicity, 877 were non-Hispanic (95.2%), 12 were Hispanic (1.3%), and 32 were Unknown (3.5%). Cancer stage at diagnosis was stage 0 in 130 patients (14.1%), stage I in 545 (59.2%), stage II in 114 (12.4%), stage III in 52 (5.6%), stage IV in 34 (3.7%), and Unknown in 46 (5.0%). During follow up, 29 patients (3.1%) had at least one documented fall and 892 (96.9%) did not.

HDS score distribution.

The HDS total score had a standard deviation (SD) of 3.6, a mean of 8.3, and a median of 8. The interquartile range (IQR) was 6 to 10, and the overall range was 0 to 19. Compared to the published categories, 873 patients (94.8%) had scores below 7, 48 (5.2%) had scores from 7 to 10, and none reached 11 to 14 or 15 or higher.

Comparison by fall status.

Patients who had at least one fall had a mean HDS score of 5.8 and a SD of 1.9. The median was 5.0, and the IQR was 4.4 to 5.6. In contrast, patients who had not had a fall had a mean HDS score of 3.8 and a median score of 4.0. The independent samples t-test and the Mann Whitney U test both indicated a difference between groups with p < 0.001. Fall incidence differed by category. Among patients with scores below 7, 13 of 873 had a fall (1.5%). Among patients with scores from 7 to 10, 16 of 48 had a fall (33.3%). No patients were in the moderate or high-risk categories. The chi square test for association between category and fall status was significant with p < 0.001.

Diagnostic performance.

At the published threshold of 10 or greater, sensitivity was 0.0% (95% confidence interval [CI] 0.0–11.7) and specificity was 100.0% (95% CI 99.6–100.0). The negative predictive value was 96.8% (95% CI 95.5–97.8). The positive predictive value could not be estimated because no true positives were identified. Treating the HDS score as a continuous predictor, the area under the curve was 0.79 (95% CI 0.70–0.88).

Threshold exploration.

At a cut score of 4, sensitivity was 96.6% and specificity was 58.0 % with negative predictive value 99.7%. At a cut score of 5, sensitivity was 75.9% (95% CI 56.2–87.4) and specificity was 77.3% (95% CI 70.2–82.5). Positive predictive value was 6.5% (95% CI 4.0–10.2), and negative predictive value was 98.9% (95% CI 97.8–99.4). At a cut score of 6, sensitivity was 51.7%, and specificity was 88.2% with negative predictive value 98.3%.

Regression and subgroup analyses.

In logistic regression, each one-point increase in HDS score was associated with higher odds of a fall (odds ratio 1.73, 95% CI 1.42–2.11, p < 0.001). The association remained after adjustment for age, race, ethnicity, cancer stage, and treatment exposures. Categorical regression was not estimated because no patients were in the moderate or high categories. Subgroup summaries by stage early versus advanced and by treatment exposure chemotherapy versus endocrine therapy did not show improved discrimination.

Discussion

This study examined the predictive performance of the HDS in patients with breast cancer. The receiver operating characteristic area under the curve was 0.79. The published cut score of 10 identified no fall events, while lower cut scores increased sensitivity with a tradeoff in specificity. Scores were higher in patients who fell, and each one-point increase in the total score was associated with higher odds of a fall after adjustment for age, race, ethnicity, cancer stage, and treatment exposures. Taken together, these results indicate that the scale carries predictive signal, but the threshold derived in neuroscience settings does not transfer to oncology care without recalibration.

Our findings align with prior work that validated the HDS in neuroscience patient populations and reported strong sensitivity at a cut score of 10, and with practice-based evaluations that showed variable performance across hospital units. The event rate in this cohort was 3%, which explains the low positive predictive values and the very high negative predictive values. In low event settings, predictive values are constrained by prevalence even when discrimination is moderate. This pattern clarifies why a threshold of 5 balanced sensitivity and specificity in this cohort and why the positive predictive value remained low despite acceptable discrimination. Our findings remained consistent among adults who were 65 years or older, supporting applicability of recalibration in older oncology populations.

This analysis has limits. It is retrospective and from a single health system, so results may reflect local practice and documentation patterns. HDS domains and fall outcomes were abstracted from electronic records rather than a nurse administered instrument, which can miss fall events, domain details, and care received outside the system. This fact may partially explain the low number of fall events, which widens confidence intervals and limits power for subgroup analyses and interaction testing. We did not assess calibration plots, reclassification, or decision curves, so clinical utility beyond sensitivity, specificity, and predictive values is not established.

Two possible short-term steps are supported if the HDS was to be used to predict fall risk in breast cancer populations with automated data collection from electronic health records. First, clinical teams may adopt a starting threshold of 5 and monitor local performance using routine data. This can include simple calibration checks, periodic review of true positives and false positives, and linkage to action steps. Second, future work should test oncology specific models that add treatment exposures, simple functional measures, and recent clinical factors such as steroid use, anemia, electrolyte disturbances, neuropathy, and assistive device use. Prospective studies with repeated measures can evaluate updated performance, external validity across services, and implementation outcomes that track action steps, resource use, and patient safety events.

Conclusion

In patients with breast cancer, the HDS shows moderate discrimination, the published threshold of 10 does not identify events, and a cut score of 5 balances sensitivity and specificity with a very high negative predictive value.