Using routine referral data for patients with knee and hip pain to improve access to specialist care

The current recommended pathway for long term conditions, including adults with knee and hip pain, is management in primary care and referral to a multi-professional assessment and treatment clinic if specialist opinion is required [1‐3]. However, there are a number of variations in the current care pathway, which represent potential inefficiency in resource use and standards of care [4] and importantly delay for patients [5]. This includes variation in where clinics are based, what profession assesses and treats patients and in the care given [5‐7]. With an aging population and rising treatment expectations the burden on healthcare resources is increasing [8]. Therefore, triage methods that streamline patients to maximise efficiency and ensure individuals receive optimal care for their needs are required. This includes ensuring early access to non-surgical treatment options such as physiotherapy, pain medication and dieticians, that treatment is given in a timely manner and in a suitable setting to meet patient needs [1, 9]. For example, an optimal outcome from a surgical consultation based on individual circumstances would be a referral for surgery, whereas sub-optimal outcome would be no definitive treatment.

To try and improve efficiency and resource use in secondary care, referral prioritisation systems have been developed for hip and or knee pain and tested to fast-track cases for surgical opinion based on general practitioner referral information [10, 11]. The limitation of these systems is that the prioritisation criteria has lacked sensitivity and specificity as individuals move between surgical and conservative pathways. The quality of the research means that insufficient evidence exists regarding which predictor variables can be used to inform decision making. Furthermore, the criterion for prioritisation rely on patient rated outcome measures and x-ray, but for conditions such as knee osteoarthritis it is recommended that decision to refer for a surgical opinion is based on discussion between patients and clinician [1]. Referral letters often contain information that underpins referral decision making, e.g. narrative description of the ways in which a given joint condition is impacting the patient’s everyday activities. However, such information has never been explored systematically in research on treatment prioritisation/ streamlining systems for orthopaedic conditions.

Analysis of free text data in general practitioner referral letters can be done using text mining techniques to create variables that can be used alongside demographic and health related data and has the potential to improve treatment prioritisation. This technique has been used successfully in the evaluation of radiology reports and health correspondence on web-based health communities and questionnaires [12‐15].

Management of these routine data from general practitioner referral letters may provide invaluable information that can predict where and by whom an individual is best seen by identifying associations between the referral information and treatment outcome [13]. Therefore, it could be used for early streamlining of the type of care an individual should receive and better resource allocation within the referral pathway. This adds to the findings of previously reported knee and or hip specific prioritisation and streamlining systems that have not included this data [8, 9].

The primary aim of the study was to identify factors from the general practitioner referral letters that can predict who would receive an optimal versus sub-optimal care pathway at the time of consultation with a specialist in an advanced physiotherapy practice, specialist GP or orthopaedic clinic. The secondary aims were to:

These findings could be used to streamline the referral process and provide recommendations in pathway re-design and streamlining patients to optimise care.

The research design was a prospective longitudinal design. All data were collected between August 2016 to January 2017 and follow-up data collection was completed in June 2018. All participants were recruited from the musculoskeletal service at one University Health Board, an administrative unit within the National Health Service. The care pathway is illustrated in Fig. 1. A consecutive sample of patients with hip and or knee pain that had been referred by their general practice for specialist opinion were screened for inclusion from orthopedic surgeon led orthopaedic clinics, specialist general practitioner clinics (specialist GP) and advanced physiotherapy practitioner clinics. The aim of the clinics was to give a specialist opinion and have a definitive treatment outcome. Individuals were eligible to take part in the study if they were referred by their general practitioner for knee and/or hip pain, if they were aged 18 or over, could provide informed consent and understand English sufficiently to be able to complete the questionnaires. The exclusion criteria were knee/hip pain secondary to other health conditions such as rheumatoid arthritis, pain secondary to knee/hip replacement, surgery for the same knee/hip within the past 12 months or already having received treatment at the primary/secondary care interface for the same condition within the previous 6 months.

A sample of 634 participants were recruited. The sample size was determined based on the volume of consecutive monthly referrals received by the service for the 6 month duration of the project. No sample size calculation was performed but we allowed for a minimum of 10 cases per potential predictor variable. This sample size also allowed for incomplete data such as missing referral letters or data on referrals. Based on 26 predictor outcomes that were explored for inclusion in the model this allowed for 23 subjects per variable on average.

At the time of attendance at their specialist consultation individuals were asked to complete a questionnaire in order to calculate knee or hip osteoarthritis outcome score (KOOS or HOOS). The KOOS and HOOS scores are validated and reliable patient rated outcome measures for assessing pain, symptoms, activities of daily living, sport and quality of life in individuals with knee and hip conditions [16].

Key data were extracted from their referral letter including age, gender, postcode, body mass index (BMI), medication, smoking status, co-morbidity data and free text referral information. From the co-morbidity data the Charlston co-morbidity index for each participant was calculated. A score of 0 indicates no co-morbidity and a score of 3 indicates high level of co-morbidity [17]. The outcome of the consultation such as the recommended treatment or further referral (e.g. consultation notes i.e. physiotherapy, surgery, discharge, follow-up, injection, imaging, dietician or orthopedic consultant referral) was extracted from the clinic outcome letter and checked by a second member of the research team.

Clinical narratives (free text) within the general practitioner referral letters were coded automatically in preparation for subsequent statistical analysis. The coding was performed against the MetaThesaurus included in the Unified Medical Language system [18]. The MetaThesaurus is a large biomedical thesaurus that is organised by concept (i.e. meaning) whose various names (or terms) are drawn from around 200 source vocabularies, e.g. SNOMED, HL7, ICD-10, DrugBank, MedDRA, etc. The free-text content of referral letters was mapped against the MetaThesaurus using MetaMap [19], a dictionary lookup application developed specifically to flexibly match biomedical terms against text and map their occurrences back to the corresponding concept identifier. We limited the lookup to concepts of certain types using the categorisation of the MetaThesaurus concepts into the Semantic Network, a taxonomy of over 130 semantic types. Specifically, we focused on the following semantic types: “Diagnostic Procedure”, “Daily or Recreational Activity”, “Disease or Syndrome”, “Finding”, “Functional Concept”, “Health Care Activity”, “Injury or Poisoning”, “Occupational Activity”, “Physiologic Function”, “Pharmacologic Substance”, “Sign or Symptom”, “Tissue”, “Therapeutic or Preventive Procedure”, “Body related concept” (an aggregation of the semantic types: “Body Location or Region”, “Body Part, Organ, or Organ Component”, “Body Space or Junction”). The MetaMap output was used to count the number of mentions per concept. This ensured that statistical analysis was based on the underlying meaning and not on surface textual representation. For example, synonyms such as ‘edema’ and ‘swelling’ would be represented by the same code. In addition, all concept mentions were also aggregated across semantic types. For example, both ‘edema’ and ‘pain’ would count toward a ‘Sign or Symptom’. Once extracted, all coded data were formatted and stored in a relational database to allow easy export for further statistical analysis.

At 6 months post the initial consultation, participants were contacted by post and asked to complete KOOS/HOOS questionnaires. The hospital database was checked to evaluate if the treatment they had been referred for had been completed, if they were still waiting or if an alternative treatment had been given.

The primary aim of this study was to identify factors predicting an optimal care pathway. An optimal care pathway was defined as one that minimises delayed treatment for the patient and results in care being delivered in a timely fashion in the right setting and by the right person. This definition was based on the literature [9, 20‐22] and by steering committee consensus who included two surgeons, a general practitioner, a physiotherapist, a member of the public, two research assistants, two medical statisticians and a computer scientist. Using this definition, a grid of optimal/ sub-optimal treatment outcomes per clinic type was created by two clinician reviewers (advanced physiotherapy practitioner and orthopaedic surgeon) independently reviewing the treatment outcome types. Agreement was achieved by consensus with a third member of the research team who was not a healthcare professional. The grid was then reviewed by the steering committee. This was then applied to the outcome data by the statistician (Table 1). Examples of optimal care from an orthopaedic clinic would be listed for surgical procedure. Non-optimal care from an orthopaedic clinic would be referral to non-surgical treatments such as physiotherapy as this would have been expected to have been done before specialist opinion in a surgeon led clinic. Optimal care from advanced physiotherapy practitioner or specialist GP may be discharge or referral for non-surgical treatment. Sub-optimal care would be an outcome of further review with no definitive outcome [1]. In the given pathway, MRI imaging and injection were not routinely available in primary care and, therefore, were rated as optimal outcomes for the advanced physiotherapy practitioner and specialist GP clinic.

All data were analysed using SPSS (version 20) and STATA (version 13). The HOOS and KOOS scores were combined as a single measure as the analysis was not specific to joint type and to manage the large volume of data based on the analysis of the individual sub scales for pain, symptoms, sports, function and quality of life. The score for each sub-scale was calculated by transforming each score to a 0–100 scale, with zero representing extreme knee problems and 100 representing no knee problems. A combined pain and function KOOS₂/HOOS₂ score was calculated using the method described by [23] to reduce the number of variables.

Baseline demographic and clinical data together with the text mining outputs from referral letters were summarised and tabulated. Frequencies were also calculated to describe important characteristics of the care pathway. A Chi square test was used to determine association between optimal pathway and clinic type, i.e. orthopaedic clinic, advanced physiotherapy practitioner or specialist GP. Logistic regression was used to investigate demographic, clinical and text concept variables associated with an optimal pathway. Variables associated at the 10% level in univariate analyses were included in a multivariate model. Variables in the multivariate model were entered using a stepwise backward selection process, with all candidate variables initially entered, and variables not significant at the 10% level sequentially removed one by one until the model could no longer be improved. A similar modelling strategy was employed for the investigation of factors predictive of KOOS/HOOS combined pain and function scores, in this case linear regression was used for continuous data. An independent T-test was used for the comparison of mean combined KOOS/HOOS score at 6 months between those on optimal and suboptimal care pathways. At 6 months post consultation predictors of combined pain and function scores were explored using a univariate analysis only.

A total of 643 participants were recruited. Referral letters were available for 586 out of 643 (91.1%) participants. The study flow chart in Fig. 2 provides details of the data included in the analysis. Based on the automatically extracted codes from the free text content of the referral letters, a total of 14 pertinent variables were identified from a set of 49 possible variables. The definition of each variable taken from the Unified Medical Langauge System [24] can be found Table 7 in Appendix.

The primary aim of this study was to identify factors from GP referral letters that can predict which patients with knee and/or hip pain would receive an optimal care pathway at the time of consultation. Factors which were found to predict an optimal care pathway were: lower BMI,, having a named disease or syndrome and taking a pharmacologic substance. Having a single diagnostic procedure predicted a sub-optimal pathway. Over 30% of participants were found not to have had an optimal care pathway. The secondary aims were to identify predictors of patient rated pain and function at time of consultation and after 6 months and to describe the characteristics of the care pathway. Variables found to predict pain and function at initial consultation were higher age, higher BMI, current smoking, with knee pain, having sign or symptoms and having therapeutic/preventative procedure and opioid medication history. Of these variables only age, BMI, smoking status and medication history were individually found to predict pain and function at 6 months post consultation. Only predictors related to BMI were predictors of both optimal care and pain and function. A key characteristic of the care pathway for individuals with knee and/ or hip pain is that treatments received varied according to the type of specialist clinic seen in.

In this study 30% of individuals did not follow an optimal care pathway which represents potential inefficiency and wasted healthcare resource. A core set of variables from the free text of referrals has been identified that should be included in a minimum information standard when referring an individual for specialist opinion for knee and hip pain. Of key importance is data on BMI as this was a predictor for both optimal care and pain and function outcomes. Patient rated outcomes for pain and function on their own were not predictors of optimal care and therefore cannot be used on their own to streamline patient referrals. A high number of patients seen in the specialist GP and advanced physiotherapy practitioner clinics received conservative treatments that could have been carried out in primary care. Finally, there was variation in the type of treatment a patient received depending on the clinic type. The recommendation from this study would be to utilize the different skill-mix of the healthcare professionals in the pathway to improve referral to conservative treatments in primary care. Further validation of a core data set at predicting optimal care to streamline referrals is required.