Simplified two-question screening for elevated Hospital Anxiety and Depression Scale (HADS) scores in orthopaedic patients: a retrospective cohort study with prospective validation

Abstract

Objectives

To assess whether it is possible to reduce the burden of completing the full Hospital Anxiety and Depression Scale (HADS) by identifying patients at risk for elevated HADS ≥11 using fewer questions.

Design

Retrospective cohort study followed by prospective validation.

Setting

A large Dutch hospital with orthopaedic patients.

Participants

13 345 orthopaedic patients with a median age of 53 years (IQR: 37–65) for the retrospective cohort; 100 patients for prospective validation.

Primary outcome measure

Incidence of HADS scores ≥11 and performance of a simplified screening model for elevated HADS.

Methods

The full HADS questionnaire consists of 14 items (7 each for anxiety and depression). Logistic regression models were developed retrospectively to predict HADS ≥11 using one or two questions. Models were evaluated based on sensitivity (>90%), negative predictive value (NPV >90%) and reduction of the response burden (>50%). The most effective model was validated prospectively in 100 patients.

Results

The retrospective cohort included 13 345 patients, with 20% (n=2671) having a HADS score ≥11. The final model, using questions 4 (‘I can laugh and see the funny side of things’) and 5 (‘Worrying thoughts go through my mind’), achieved an area under the curve of 0.934, sensitivity of 95%, NPV of 98% and reduced the questionnaire burden by 57%. Prospective validation confirmed high sensitivity (94%) and NPV (98%) for identifying patients at risk of elevated HADS scores.

Conclusions

In a large Dutch hospital, 20% of orthopaedic patients exhibited elevated HADS scores (≥11). Screening for elevated HADS using only two questions was highly effective, with significant reductions in patient burden. Patients who screen positive should complete the full HADS for accurate diagnosis and management. This approach could reduce the need for full HADS completion in over half of the patients, without sacrificing screening accuracy.

STRENGTH AND LIMITATIONS OF THIS STUDY.

• This is the largest study investigating Hospital Anxiety and Depression Scale (HADS) scores in the orthopaedic population, and therefore gives reliable insights in the current prevalence of high HADS scores.

• Screening for a HADS ≥11 with only two questions is possible in an orthopaedic population

• Selection bias was prevented as every patient at the outpatient clinic had to fulfil the HADS questionnaire.

• Validation was performed on anonymous data, this makes comparison of baseline characteristics not possible.

Introduction

Each year, thousands of patients undergo elective orthopaedic procedures such as total knee arthroplasty or total hip arthroplasty, that are largely beneficial.¹ However, some patients are dissatisfied after these procedures due to ongoing pain or reduced function.2,4 While some patients have clear reasons for their dissatisfaction, such as prosthetic joint infection, hardware failure or aseptic loosening of the implant, others do not. Patients may have persistent pain and functional limitations despite appropriate treatment,5,7 with proportions reported as high as 20%–30%.^{8 9} The role of psychological factors in postoperative pain and reduced function remains relatively understudied. Some previous research has, however, indicated their potential significance.^{10 11} Patients with a history of anxiety or depression, for instance, are more prone to experience postoperative pain and reduced function,^{10 12 13} and preoperative depression and anxiety are significant predictors of poor outcomes.¹⁰ Notably, a substantial proportion of patients with osteoarthritis (41%) exhibit clinically significant anxiety and/or depression.¹⁴ Consequently, screening for anxiety or depression among patients with orthopaedic issues is paramount, as these patients may benefit from preoperative mental health interventions.¹⁵ Recognising psychological symptoms is challenging for healthcare providers in non-psychiatric settings due to little mental health training and time constraints.¹⁶ The Hospital Anxiety and Depression Scale (HADS) is a widely used self-reported screening questionnaire that consists of fourteen questions and has been validated for patients with physical illnesses, including orthopaedic patients.¹⁷ Nevertheless, the requirement to complete this questionnaire imposes an administrative burden on patients in our clinic, as they repeatedly stated to our orthopaedic surgeons. Given the importance of identifying psychological comorbidities, this study investigated the feasibility to reduce the burden of completing the full HADS by identifying patients at risk for an elevated HADS (≥ 11) using less questions.

Patients/methods

Study design and setting

This is a single centre study conducted at the Department of Orthopedic Surgery of a large teaching hospital (Onze Lieve Vrouwe Gasthuis (OLVG)) in Amsterdam, the Netherlands. The study received approval from the institutional review board (WO 22.205) and informed consent was waived for the retrospective investigation. For the prospective part, patients anonymously signed the patient information form. No patients or members of the public were involved in the design, conduct, or reporting of this study. The main goal of this study was to develop and validate a brief screening rule to identify patients at risk of elevated HADS scores among orthopaedic patients, using fewer HADS questions, without compromising its screening efficacy. To this end, a stepped approach was employed, focusing on achieving high sensitivity and a high negative predictive value (NPV) in the preliminary phase, thereby minimising the necessity for patients to complete the full HADS questionnaire.

This study consists of a two-part investigation: initially, a retrospective analysis of existing HADS questionnaire data to identify the most effective questions for achieving a high NPV, followed by a prospective phase to validate the efficacy of these selected questions.

In the first part, retrospective data from new patients visiting the outpatient clinic between December 2020 and December 2022 were used. The data was extracted from the digital intake, a process where each new orthopaedic patient at our hospital receives a digital questionnaire that not only addresses routine questions related to their specific orthopaedic complaints but also includes the HADS questionnaire.¹⁸ The questionnaires had to be fulfilled before intake took place. Patients who visited the clinic for emergency appointments (such as suspected septic arthritis) and patients under 18 years of age were excluded since they did not complete the digital intake. Patients with fractures had their first appointments at the trauma surgery outpatient clinic, and were therefore not included in the current study (in this hospital, the trauma surgeon treats 80% of fractures, and the orthopaedic surgeon 20%). The data from the HADS were used to develop models consisting of 1 or 2 questions predicting scores of 11 or higher according to predefined criteria for success.¹⁷

In the second, prospective part, the best performing model was validated at the outpatient clinic between June 2023 and September 2023. All patients attending a scheduled appointment at the orthopaedic outpatient clinic between June and September 2023 were approached for participation. Patients were eligible if they were 18 years or older and sufficiently proficient in Dutch or English to complete the questionnaire independently. All eligible patients were consecutively invited to participate. Those who provided consent completed the short version (two selected questions), followed by the remaining HADS to allow calculation of the final HADS scores. The remaining questions were placed on the back of the page to ensure that participants’ responses to the two key questions on the front were not influenced by the other items. The outpatient clinic assistant personally explained this requirement to ensure full completion. A sample size of 100 patients for the prospective validation was chosen in line with the COSMIN (COnsensus-based Standards for the selection of health Measurement INstruments) guideline, which considers a sample size of at least 100 participants as ‘excellent’ for studies assessing measurement properties.¹⁸

COnsensus-based Standards for the selection of health Measurement INstruments

guideline, which considers a sample size of at least 100 participants as ‘excellent’ for studies assessing measurement properties.¹⁸

This study adhered to the STROBE (STrengthening the Reporting of OBservational studies in Epidemiology) guidelines for observational studies.¹⁹

Model requirements and selection 

The HADS questionnaire consists of 14 questions, with 7 items for anxiety and 7 for depression (figure 1). The scoring for each question ranges from zero to three. A HADS cut-off of 11 or higher was used to indicate a probable case of anxiety (HADS-A) or depression (HADS-D), consistent with established cut-offs for the domain rather than the total HADS score. This value has been shown to be appropriate for identifying probable cases of anxiety and depression with high specificity and reasonable sensitivity.¹⁷

Figure 1. English Hospital Anxiety and Depression Scale (HADS) questionnaire.

Several models were developed (details in the Statistics section), which had to fulfil the following criteria for an ideal screening test. First, the models had to achieve the highest possible sensitivity (above 90%) to avoid false-negative results (patients should not be missed during screening). Second, it was aimed to achieve a high NPV (above 90%), so that patients who receive a negative test result truly have a negative HADS score. Thirdly, the model should reduce the number of patients who need to complete the full questionnaire by at least 50%. Finally, in case models with more than one question will be selected, the preference is to include a question for both depression and anxiety. A model with an area under the curve (AUC) above 0.80 was considered as having good discriminative ability and an accuracy of 80% was considered as being precise. Models meeting the first three criteria (sensitivity, NPV and response burden reduction) were considered effective. Among the effective models, the primary ranking metric was the AUC, with the model showing the highest AUC considered superior. In cases where multiple models had similar AUC, sensitivity was used as a secondary criterion to determine the best-performing model. In cases where multiple models had comparable performance, consensus decision-making among the authors (van de Kuit, van Maarseveen and Kempen) was used to select the final model for prospective validation. Questions that represented both domains of the HADS, selecting one item for anxiety and one for depression to ensure content validity in addition to statistical performance, were prioritised.

Statistics

Descriptive statistics of the HADS were calculated as mean±SD or median (IQR), depending on normality (Kolmogorov-Smirnov test). Patients with >20% missing data were excluded.

Model development

Binary logistic regression analysis with outcome HADS ≥11 was performed. First, models with only one question were created (univariate logistic regression, with HADS as categorical variable). OR’s per question were calculated. Subsequently, models with 2 questions were created, comprising a total of 91 logistic regression models of 2 questions (eg, 1+2, 1+3 … 13+14). Models including three questions (or more) were developed in a similar manner, iteratively adding questions until satisfactory model performance was achieved. For each of the models, performance measures were calculated (AUC, sensitivity, specificity, Negative Predicitive Value (NPV), Positive Predictive Value (PPV), accuracy). Models were reviewed based on the earlier described requirements.

Threshold/cut-off determination

In classification regression analysis, the default threshold for interpreting probabilities is 0.5. If the calculated probability is below the cut-off, the probability is assigned to be negative, and values greater than or equal to the cut-off are assigned to be positive. Each different cut-off value belongs to different performance measures (the lower the threshold, the easier a ‘positive’ test result). Threshold adjustment therefore gives possibilities in optimising the model’s performances to our requirements. The best models were selected based on their AUC and sensitivity. After this, model optimisation was achieved through adjusting their threshold levels. Applying threshold shifting to all models was deliberately avoided, as this could lead to excessive data-driven decision-making and reduce reproducibility. This was performed by plotting the Receiver Operating Characteristic (ROC) curve with its coordinates. In this overview, all cut-off points with corresponding sensitivity and 1-specificity are displayed. The best cut-off value was then selected on the following criteria: sensitivity as high as possible, with specificity above 60%, as we prioritised sensitivity to minimise false-negative results in the screening setting. Then, logistic regression analysis was repeated with the new cut-off value, resulting in new performance measures (the final performance measures).

Analysis of the prospective part (validation)

Logistic regression analysis was repeated and performance measures were calculated with the cut-off as calculated in the retrospective part. In case good performances are achieved in the prospective part, the model will be translated to clinical practice.

Translation to clinical practice

In this part, the statistical cut-offs will be translated to cut-off values for the questions that can be used in clinical practice. Herewith, it will be determined which patients do and do not have to complete the complete questionnaire. The cut-off values for the questions are calculated with the following logistic regression formula:

$${\displaystyle {\displaystyle p=\frac{1}{\left(1+e^{-\left({\beta }_0+{\beta }_{1,1}•x_{1,1}+{\beta }_{1,2}•x_{1,2}+{\beta }_{1,3}•x_{1,3}+{\beta }_{2,1}•x_{2,1}+{\beta }_{2,2}•x_{2,2}+{\beta }_{2,3}•x_{2,3}\right)}\right)}}}$$

,in which p represents the calculated probability, ${\beta }_0$ the intercept of the model, ${\beta }_{\mathrm{1,1}}$ the coefficient of $x_{\mathrm{1,1}}$ if question 1 equals 1, ${\beta }_{\mathrm{1,2}}$ the coefficient of $x_{\mathrm{1,2}}$ if question 1 equals 2, ${\beta }_{\mathrm{1,3}}$ the coefficient of $x_{\mathrm{1,3}}$ if question 1 equals 3 and then the same for the other questions. $x_{\mathrm{1,1}}$ is only 1 if the value of question 1 equals 1, $x_{\mathrm{1,2}}$ is only 1 if the value of question 1 equals 2 and so forth.

If the cut-off value is calculated, and the intercept and coefficients are calculated in the logistic regression analysis, cut-off values for the questions can be calculated. Statistics were performed in RStudio, V.1.4.1106 and IBM SPSS Statistics V.29.0.

Results

Incidence of increased HADS

A total of 13 374 patients were eligible for inclusion, of which 29 patients (0.2%) had missing data and were excluded, leaving 13 345 patients for analysis. Median age was 53 years (IQR: 37–65) and 55% of patients were female. HADS score was negatively skewed, and median score was 9 (IQR: 5–16). Of these patients, 2671 had an increased (≥11) HADS for depression and/or anxiety (20%). Considering the subdomains, 1895 (14%) patients had an increased HADS (≥11) for anxiety and 1842 (14%) had an increased HADS for depression.

Model development

First, binary logistic regression with only one HADS question was performed, which showed fourteen models with an AUC between 0.769 and 0.870 (table 1), sensitivity ranging from 0.200 to 0.781, specificity ranging from 88.0% to 99.2%, NPV ranging from 82.3% to 98.0%, PPV ranging from 52.0% to89.2% and accuracy ranging from 0.817 to 0.863. Since none of these models did meet our predefined thresholds, 91 logistic regression models consisting of two questions were developed (eg, 1+2, 1+3 … 13+14). The AUC of the models ranged from 0.861 to 0.937 (supplementary appendix, online supplemental Table S1 (table 2). To further explore whether model performance would meaningfully improve with additional items, we did evaluate three-item combinations. These models showed only minimal increases in AUC, up to 0.022 at most. The model consisting of questions 4 and 5 had the highest sensitivity and was, after consensus-decision making, selected for threshold shifting, and hereafter named as ‘Model 1’. Model 1 consists of question 4 ‘I can laugh and see the funny side of things’ (a question for depression) and question 5 ‘Worrying thoughts go through my mind’ (a question for anxiety). The model showed (at the default cut-off of 0.50) an AUC of 0.934, a sensitivity of 79.8%, specificity of 91.2%, PPV of 96.3%, a NPV of 94.7% and an accuracy of 88.9%. After threshold shifting (online supplemental data, appendix 2, online supplemental figure S1 and table S2), a cut-off of 0.082 was selected, which resulted in a sensitivity of 95.4%, specificity of 69.9%, a PPV of 44.2%, a NPV of 98.4% and an accuracy of 75%. In total, 5766 of 13 345 (43%) had a positive screening. Consequently, 7579 (57%) of the patients did not have to proceed filling in the complete HADS afterwards after screening with two questions. Of the 10 674 patients with a HADS <11, the burden of completing the full HADS could be decreased in 7579 patients (71%).

Table 1. Univariate logistic regression for all 14 questions of the HADS questionnaire with ORs.

Model (n=1)	A/D	OR (95% CI) per score (1, 2 and 3; 0 is index)	AUC	Sensitivity	Specificity	PPV	NPV	Accuracy
Question 1	A	1: 3.395 (2.747 to 4.196) 2: 37.509 (30.315 to 46.410) 3: 137.309 (105.960 to 177.932)	0.837	0.714	0.889	0.618	0.925	0.854
Question 2	D	1: 5.325 (4.312 to 6.576) 2: 29.657 (24.163 to 36.400) 3: 187.161 (143.179 to 244.653) Onderkant formulier	0.830	0.231	0.987	0.820	0.980	0.836
Question 3	A	1: 4.391 (3.797 to 5.079) 2: 23.611 (20.542 to 27.138) 3: 158.872 (125.835 to 200.583)	0.853	0.688	0.898	0.628	0.919	0.856
Question 4	D	1: 9.694 (8.501 to 11.055) 2: 102.319 (85.262 to 122.788) 3: 50.282 (37.592 to 67.256)	0.841	0.431	0.971	0.791	0.871	0.863
Question 5	A	1: 3.636 (2.952 to 4.478) 2: 43.907 (35.902 to 53.698) 3: 335.790 (250.621 to 449.901)	0.870	0.781	0.880	0.520	0.941	0.860
Question 6	D	1: 11.830 (10.236 to 13.672) 2: 41.631 (35.325 to 49.063) 3: 34.696 (28.753 to 41.868)	0.824	0.488	0.912	0.581	0.870	0.827
Question 7	A	1: 5.589 (4.535 to 6.887) 2: 36.083 (29.771 to 43.733) 3: 109.790 (86.436 to 139.453)	0.843	0.200	0.980	0.712	0.830	0.824
Question 8	D	1: 4.182 (3.066 to 5.704) 2: 33.344 (24.628 to 45.145) 3: 107.431 (78.429 to 147.158)	0.829	0.350	0.955	0.662	0.823	0.834
Question 9	A	1: 6.379 (5.714 to 7.121) 2: 52.638 (43.587 to 63.567) 3: 106.209 (73.802 to 152.845)	0.798	0.332	0.981	0.811	0.980	0.850
Question 10	D	1: 4.751 (4.269 to 5.287) 2: 26.169 (22.384 to 30.593) 3: 20.439 (16.538 to 25.260)	0.769	0.357	0.957	0.675	0.856	0.837
Question 11	A	1: 3.806 (3.345 to 4.330) 2: 12.068 (10.518 to 13.847) 3: 26.661 (22.406 to 31.724)	0.769	0.203	0.970	0.632	0.892	0.817
Question 12	D	1: 6.273 (5.499 to 7.156) 2: 41.191 (35.567 to 47.703) 3: 36.547 (29.630 to 45.078)	0.829	0.550	0.929	0.659	0.891	0.853
Question 13	A	1: 7.865 (7.064 to 8.775) 2: 91.495 (70.558 to 118.594) 3: 253.000 (149.343 to 428.06)	0.803	0.272	0.992	0.892	0.845	0.846
Question 14	D	1: 5.160 (4.497 to 5.921) 2: 20.936 (18.361 to 23.873) 3: 56.814 (46.318 to 69.688)	0.825	0.638	0.880	0.571	0.907	0.831

AUC, area under the curve; HADS, Hospital Anxiety and Depression Scale; NPV, negative predictive value.

Table 2. The 5 best models in terms of sensitivity from the 96 developed models (standard cut-off value 0.5).

Model	AUC	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)
Questions 1 and 4	0.925	71.3	92.7	70.9	92.8
Questions 5 and 13	0.891	71.3	93.6	73.6	92.9
Questions 3 and 4	0.886	71.6	92.8	71.4	92.8
Questions 5 and 6	0.882	74.1	91.8	69.3	93.3
Questions 4 and 5	0.889	79.8	91.2	96.3	94.7

AUC, area under the curve; HADS, Hospital Anxiety and Depression Scale; NPV, negative predictive value.

Validation of models

A total of 100 patients were included for the prospective validation study of the selected model. In total, four patients had some missing data (forgot to fill in one question) and were excluded for analysis. The median HADS-A score was 4 (IQR 3–7). The median HADS-D was 4 (IQR 2–6). In total, 18 (19%) patients had an increased HADS for anxiety and/or depression.

Logistic regression for Model 1 showed an AUC of 0.925 with a sensitivity of 78%, specificity of 95%, PPV of 78% and a NPV of 95%, for a model cut-off of 0.50. After adjusting the threshold to a cut-off of 0.082, performances result in a sensitivity of 94%, specificity of 73%, PPV of 44% and NPV of 98%.

Translation to clinical practice

In the development part, a cut-off of 0.082 was selected. The following scores per question lead to a $p$ above 0.082 and therefore a positive test (online supplemental data, appendix 3):

• Question 4≥2;

• Question 5≥2; and

• Question 4≥1 AND question 5≥1.

These patients should complete the full HADS questionnaire. For the rest, the full HADS can be abandoned after completion of the two questions (figure 2).

Figure 2. Flowchart for use of the new Hospital Anxiety and Depression Scale (HADS) model.

Discussion

This study showed that screening for a HADS score ≥11 with only two of its’ questions is possible in an orthopaedic population, with a high sensitivity, high NPV and a reduction in 57% of patients that have to complete the 14-number questionnaire. Even more, a reduction of 71% is reached for patients with a HADS <11. This finding contributes to lowering the response burden in patients, without losing screening capacity for elevated HADS. Patients who screen positive should complete the full HADS for accurate diagnosis and management.

To our knowledge, no previous studies investigated whether screening for an increased HADS is feasible. This is not remarkable, considering that the HADS was already developed as a ‘short’ questionnaire for a general medical population of patients.¹⁷ But, one study investigated whether the question ‘Are you depressed?’ is performing less than screening for depression with an increased HADS-D. The study showed that in testicular cancer survivors who answered ‘No’ on the question ‘Are you depressed?’, only 0.05% had an increased HADS-D.²⁰ This result is in line with our study that suggests that limiting the response burden may still result in adequate screening. Comparing performance measures and cut-off values is not possible because of lacking data.

The HADS questionnaire was validated on a cut-off value of 11 for probable depression or anxiety disorder, and the questionnaire handles a cut-off value of 7 for possible depression or anxiety disorder.¹⁷ In this study the cut-off of 11 (for one of the domains) was maintained. In the general medical population, the sensitivities and specificities of the HADS are known to vary between 0.44–0.75 and 0.92–0.95, respectively, for the cut-off value of 11. For a lower cut-off (eg, 8 or higher), the sensitivity is a bit higher (0.74–0.82).^{21 22} This should be kept in mind as using the HADS questionnaire misses false-negative patients itself already.

Question 4 ‘I can laugh and see the funny side of things’ and question 5 ‘Worrying thoughts go through my mind’ were in this study selected for the final model development because they fitted well within the digital intake about the orthopaedic complaints. These are accessible questions to ask (orally or digitally) patients, facilitating implementation in clinical practice. Now, in our hospitals, the 14-point HADS questionnaire in the digital intake will be replaced by the two screening questions to implement the findings of this study in clinical practice, according to figure 2. Although items 4 and 5 were selected in this study, the findings should not be interpreted as a definitive recommendation for these specific items; rather, our results demonstrate that elevated HADS scores can be accurately predicted using only two items, and alternative item combinations may yield comparable performance in other settings.

In future studies, (e-)interventions could be used for patients with an increased HADS. Providing help to patients willing to participate might reduce psychiatric symptoms peri-operatively (such as delirium²³) and optimise postoperative orthopaedic outcomes. These e-health interventions offer accessible and convenient ways for orthopaedic patients to receive support, which can be particularly valuable in the peri-operative period when patients may experience heightened stress and anxiety. By incorporating mental e-health interventions into peri-operative care, healthcare providers can address not only the physical aspects of surgery but also the psychological well-being of patients. This holistic approach to care has the potential to improve overall patient outcomes and enhance the patient experience during the peri-operative period and beyond.²⁴ Furthermore, a meta-analysis already showed improved outcomes (reduced pain, reduced perioperative anxiety and improved recovery) in a healthy population with e-learnings and psychological interventions.^{11 15} Future studies could explore the effectiveness of such e-health interventions in optimising orthopaedic outcomes.

The strengths and limitations should be taken into consideration when interpreting the results. The first strength is that this is, to our knowledge, the largest study with HADS scores within this field, which gives reliable insight in the current prevalence of increased HADS in the orthopaedic population. Second, this study avoids selection bias. Selection bias may arise when participants self-select into a study, potentially leading to distorted findings, as individuals with certain characteristics (eg, not being depressed), may be more inclined to participate. In this digital intake process, all patients visiting the orthopaedic outpatient clinic (being representative of the population and thus minimising the risk of selection bias) are systematically asked to complete the questionnaire prior to their initial consultation, thereby minimising the risk of selection bias.²⁵

This study also has some limitations. First of all, this was a retrospective analysis of prospectively collected data, which resulted in some missing data, for which some participants were excluded. The numbers were extremely low: out of 13 374 eligible patients, only 29 patients (0.2%) were excluded due to >20% missing responses. Additionally, four patients had a single missing item in the prospective study and were therefore excluded. These missing responses appear to be due to patients inadvertently skipping questions, and therefore we consider the missing data to be reasonably missing at random, and the exclusion of these few patients is unlikely to have introduced bias in our analyses. Second, data of the prospective part was processed anonymously. Validation was performed without comparison of baseline characteristics. This lack of baseline comparison may limit the ability to fully assess the effectiveness of the screening approach. Third, a psychometric limitation of this study is that there is no established ‘core question’ within the HADS; all 14 items contribute to measuring the constructs of anxiety and depression. While psychometric studies^{26 27} suggest that individual items have reasonable representativeness for their respective subscales, reducing the scale to two items inevitably limits construct and content coverage. However, the aim of this study was not to reassess the construct or content validity of the HADS, but to evaluate whether elevated HADS scores can be accurately identified using a minimal number of items for screening purposes. Consequently, any positive findings should always be followed up with completion of the full HADS questionnaire. Fourth, although the logistic regression cut-offs enable practical implementation of the model in clinical practice, they do not provide full evidence of predictive validity. Formal comparison with the complete HADS would be needed in future studies to confirm equivalence and predictive performance. Fifth, our objective was not to establish statistically significant differences between models (eg, with DeLong test), but to identify one meeting predefined clinical performance thresholds (sensitivity >90%, NPV >90% and ≥50% reduction in response burden). Given the minor performance variations and consistent analytic approach, we chose a pragmatic model selection strategy rather than formal statistical testing, as this better reflects real-world clinical application. Finally, the models predict elevated HADS scores (≥11) rather than fully capturing the underlying construct required for a clinical diagnosis. Therefore, these individuals should be referred for further assessment by a qualified mental health professional to establish a clinical diagnosis and determine appropriate follow-up care.

Future studies could include prospective evaluation of multiple models, for example through an A/B testing design, to further strengthen the evidence for clinical applicability and reduce order bias.²⁸ Also, we want to emphasise that other combinations of questions could be used in clinical practice, but these models were not validated in this study.

To conclude, this study showed that screening for a HADS score ≥11 is possible with two questions with a very high sensitivity and NPV in an orthopaedic patient population. This could reduce the burden to complete the full HADS for 57% of the patients, and 71% for patients with a HADS <11.

This means that a large proportion of patients may be spared from completing the entire 14-item HADS questionnaire while still achieving effective screening for anxiety and depression symptoms. It is recommended that patients with a score above the models’ cut-off should still complete the full questionnaire for adequate screening on symptoms of anxiety and depression.

Data availability statement

Data are available upon reasonable request.