Pointers to earlier diagnosis of endometriosis: a nested case-control study using primary care electronic health records

Endometriosis is a common gynaecological condition in which there is often a long time between first primary care consultation and diagnosis. 1–4 A longer time to diagnosis is associated with prolonged symptoms, particularly pain and 5,6 subfertility, along with patient frustration and demoralisation. 7 Endometriosis can be difficult to diagnose clinically; its symptoms are both common 8 and non-specific, so are often considered by GPs as part of the normal menstrual experience, 9 or attributed to other conditions. 5 The use of very detailed questions about symptoms can increase diagnostic accuracy. 10 However, current biomarkers 11 and imaging 12 have limited benefit, and there is substantial variation in guideline recommendations for diagnosis and management of this condition. 13 Most research on the clinical features of endometriosis in primary care has focused on features present at a single point in time, typically the time of diagnosis. 5,14 However, with endometriosis, the symptoms at any single point in time have only limited predictive value 2 and the problem of delays in diagnosis requires an understanding of when symptoms first appear . Although data in electronic records contain many single items, experienced practitioners typically recognise composite patterns that involve combinations of items. For example, repeated episodes of dysmenorrhoea, except when taking hormonal contraception, 15 are recognised by experienced clinicians as having diagnostic value in endometriosis. Although such knowledge-derived features 16 are not immediately present in electronic records, they can be constructed. 17 However, the authors are not aware of studies that have attempted to do this using primary care data or for endometriosis. This study aimed to: (a) construct enriched datasets from electronic health records, which contained conventional and composite features potentially predictive of endometriosis; (b) examine the association of these features with a subsequent diagnosis of endometriosis in a nested case-control study; and (c) examine the relationship of these features to diagnosis at different time periods before the date of diagnosis.

Data source Data from the Practice Team Information (PTI) database, a subset of the Primary Care Clinical Informatics Unit Research database held by the University of Aberdeen, were obtained. It includes anonymised data from primary care electronic health records of approximately 224 000 patients registered with a primary care physician, and is broadly representative of the Scottish population Research C Burton, MD, professor of primary medical care, Academic Unit of Primary Medical Care, University of Sheffield, Sheffield; Institute of Applied Health Sciences, University of Aberdeen, Aberdeen. D Ayansina, MBBS, research fellow; S Bhattacharya, PhD, senior lecturer, Institute of Applied Health Sciences; L Iverson, PhD, research fellow, Institute of Applied Health Sciences; D Sleeman, PhD, emeritus professor, Computing Sciences, Natural and Computing Sciences, University of Aberdeen, Aberdeen. L Saraswat, PhD, consultant gynaecologist, Aberdeen Royal Infirmary, Aberdeen. Address for correspondence Christopher Burton, Academic Unit of Primary Medical Care, University of Sheffield, Samuel Fox House, Sheffield, S5 7AU, UK. E-mail: [email protected] Submitted: 16 May 2017; Editor’s response: 9 June 2017; final acceptance: 11 August 2017. ©British Journal of General Practice This is the full-length article (published online 7 Nov 2017) of an abridged version published in print. Cite this version as: Br J Gen Pract 2017; DOI: https://doi.org/10.3399/bjgp17X693497 Christopher Burton, Lisa Iversen, Sohinee Bhattacharya, Dolapo Ayansina, Lucky Saraswat and Derek Sleeman Pointers to earlier diagnosis of endometriosis: a nested case-control study using primary care electronic health records

Background Endometriosis is a condition with relatively non- specific symptoms, and in some cases a long time elapses from first-symptom presentation to diagnosis. Aim To develop and test new composite pointers to a diagnosis of endometriosis in primary care electronic records. Design and setting This is a nested case-control study of 366 cases using the Practice Team Information database of anonymised primary care electronic health records from Scotland. Data were analysed from 366 cases of endometriosis between 1994 and 2010, and two sets of age and GP practice matched controls: (a) 1453 randomly selected females and (b) 610 females whose records contained codes indicating consultation for gynaecological symptoms.

Composite pointers comprised patterns of symptoms, prescribing, or investigations, in combination or over time. Conditional logistic regression was used to examine the presence of both new and established pointers during the 3 years before diagnosis of endometriosis and to identify time of appearance.

A number of composite pointers that were strongly predictive of endometriosis were observed. These included pain and menstrual symptoms occurring within the same year (odds ratio [OR] 6.5, 95% confidence interval [CI] = 3.9 to 10.6), and lower gastrointestinal symptoms occurring within 90 days of gynaecological pain (OR 6.1, 95% CI = 3.6 to 10.6). Although the association of infertility with endometriosis was only detectable in the year before diagnosis, several pain-related features were associated with endometriosis several years earlier .

Useful composite pointers to a diagnosis of endometriosis in GP records were identified. Some of these were present several years before the diagnosis and may be valuable targets for diagnostic support systems.

diagnosis; electronic health records; endometriosis; primary care.e816 British Journal of General Practice, December 2017 with regards to age, sex, deprivation, and urban/rural ratio mix. It includes data collected annually between 2004 and 2010. Practices in the PTI project were expected to record every clinical encounter using Read Codes for clinical diagnoses and/ or main reasons for consultation. All GP prescriptions were automatically recorded. Investigations and therapeutic procedures were coded differently over time — increasing towards the end of the database period. Populations This study was a nested case-control study. Cases were females with a diagnosis of endometriosis, who were born after 1 January 1974 and were, therefore, ≤36 years on 1 January 2010. This enabled us to capture teenage menstrual symptoms for the majority of females and avoid the possibility that an apparent new diagnosis in an older female was actually a historical diagnosis being recorded for the first time due to the creation of computerised record summaries. Population controls were randomly selected for each case and individually matched by age and GP practice, with up to four controls per case (subject to availability). A second control group comprised females with codes for gynaecological symptoms (pain, menstrual symptoms, or infertility) but with no recorded diagnosis of endometriosis. These controls were also randomly selected for each case and individually matched by age and GP practice, with up to four symptomatic controls per case. The index date for cases was defined as the date of diagnosis of endometriosis and for controls as the date of diagnosis of endometriosis in the matched case. All cases and controls were required to have been registered with their GP practice for at least 1 year before the index date. Data extraction and preparation Box 1 lists the key data extracted and the categories into which related items were grouped. Most items were allocated to a single time point. However, for contraception prescriptions, which commonly lasted for 6 months or longer, details were used about each prescription to estimate the onset and offset of contraception using

previously employed to ascertain the continuity of prescribing. 18 The data was enriched by introducing composite features that were based on the clinical experience of the investigators and on interviews with 10 experts (six gynaecologists, two specialists in reproductive health, and two representatives of a lay support organisation). Interviews sought to identify tacit patterns in symptoms, which clinicians thought may be predictive of a diagnosis, were audio recorded, transcribed, and analysed thematically. Composite features were specified according to one of five relationships: proximity, following, separated, during, and exclusive. These are summarised in Box 2. The presence of each feature (single and composite) was ascertained in the record of each individual at any time, and during a How this fits in Endometriosis is a relatively common condition but the time from first presentation to diagnosis is often longer than ideal as symptoms are non- specific. This study used anonymised GP record data to construct new pointers to diagnosis, which identified patterns of symptoms in time. Distinct episodes of gynaecological pain and combinations of gynaecological pain on one occasion with menstrual symptoms or lower gastrointestinal symptoms on another appear to be useful pointers to endometriosis. Patterns such as these make sense to clinicians and could be integrated into electronic diagnostic support systems. Box 1. Categories of data grouped by data type Data type Data description Included data categories Specific features Classical features of Pain (pelvic pain, dyspareunia, dysmenorrhoea) endometriosis (pelvic pain, Menstrual (flow) dysmenorrhoea, dyspareunia Infertility and infertility) 2,5,9,14 Ovarian (for example, cysts) Non-specific Abdominal pain and Menstrual (timing) symptoms gastrointestinal symptoms, Genital/other gynaecological fatigue, urinary symptoms; Urinary additional diagnoses, including Lower GI irritable bowel syndrome 5 Upper GI Fatigue Diagnostic tests Primary care tests, referred Full blood count and procedures investigations such as diagnostic Genital swabs ultrasound, and specialist Laparoscopy procedures such as laparoscopy Abdominal or pelvic ultrasound Thyroid function Treatments Hormonal treatment for Hormonal treatment endometriosis (for example, Contraception gonadotropin-releasing NSAID hormone agonists) Codeine or other opioids Prescriptions for contraception Tricyclic Analgesic drugs SSRI and related antidepressants Antidepressant drugs Lower GI = pain, bloating, irritable bowel syndrome. NSAID = non-steroidal anti-inflammatory drugs. SSRI = selective serotonin reuptake inhibitor . Upper GI = dyspepsia, reflux, nausea. British Journal of General Practice, December 2017 e817 series of overlapping 3-year time windows set at different intervals from the index date (for diagnosis or matching). The windows were defined using intervals between the end of the window and the index date of 0, 3, 6, 12, 18, 24, and 36 months. The appearance of statistical associations between available information in the record and diagnosis over time were examined by comparing the same measure in different windows. The purpose of this was to differentiate between features that were present long before diagnosis (and may thus indicate missed diagnostic opportunities) and those that appeared only shortly before diagnosis (and may thus have triggered referral). Analysis of association of features and patterns with diagnosis Conditional logistic regression was carried out to examine the association between each feature (conventional or composite) and the diagnosis of endometriosis. Each feature was reported as either present or absent within the time period. Rather than use counts of how often a feature occurred, the ‘separated’ composite variables were used to indicate multiple episodes. Conditional logistic regression was conducted for all features for which at least 10 individuals (cases or controls) had the feature present and reported as the odds ratio (OR), with 95% confidence intervals (CIs). All analyses were conducted in R 3.3.2 (version 2016). The analysis was conducted separately with population and symptomatic control groups. For the population comparison all cases and their matched controls were included. For the symptomatic comparison, only cases that had recorded symptoms and their matched controls were included. For the time window analysis, the data were limited to females who had been registered with their practice for at least 1 year before the beginning of the gap. The odds ratios for each feature at each of the six different time gaps were plotted in order to visualise the appearance of predictive features over time.

Patient characteristics Data from 366 cases and 1453 matched population controls were obtained. Of these, 243 cases had gynaecological symptoms (pain, menstrual symptoms, and infertility) and were matched to a further 610 controls with comparable symptoms. The median age at diagnosis was 25 years, interquartile range 22–28 years, and age at diagnosis was <20 years in 47 (12.8%) cases. Data quality In total, 191 cases (52.2%) were registered with the same GP practice for at least 5 years before diagnosis and, therefore, had continuous records in the PTI database; 114 (31.2%) were registered for at least 8 years before diagnosis. Similar proportions were seen for population controls (746 [51.3%] and 469 [32.3%] respectively), but more of the symptomatic controls had been registered for these time periods (414/610, 67.9% and 273/610, 44.8%). A recorded code for laparoscopy was found in only 47 (12.8%) cases despite this being the commonest diagnostic procedure for endometriosis. This is likely to represent a preference for recording the diagnosis rather than the procedure by which it was made, although instances of a clinical diagnosis being entered without any confirmatory tests cannot be excluded. Likewise, there were few coded surgical procedures, for example, 13 cases (3.5%) had a recorded operation for tubal or ovarian problems excluding diagnostic laparoscopy. These procedures were excluded from the analysis, focusing instead on clinical features, investigations, and medical treatments. Occurrence of diagnostic features There were 145 cases (39.6%) that had a code recorded for gynaecological pain (dysmenorrhoea, pelvic pain) during the 3 years prior to diagnosis and 39 (10.7%) had a code for infertility. And 198 cases (54.1%) had neither of these during the 3 years prior to diagnosis. The numbers and proportions of females with at least one instance of each feature, Box 2. Types of composite features used in constructing predictors Relationship Specification Example Proximity An occurrence of one feature within a given Pain and fatigue within 90 days of number of days of the other but with no each other specification of which should come first Following An occurrence of one feature within Pain occurring within 90 days of a given number of days of the other with estimated cessation of contraception specification of which should come first Separated Two consecutive recordings of a Two consecutive episodes of pain single feature occurring at least a separated by at least 180 days given number of days apart (this permits differentiation of separate episodes from repeated consultation during the same episode) During An occurrence of a symptom or other feature Pain during estimated duration after the onset, and before the expected of prescription for contraception offset, of a contraception prescription Exclusive A feature only occurring in the absence Pain but only outside of estimated of another periods of prescribed contraception e818 British Journal of General Practice, December 2017 either in the 3 years prior to the index date or at any time, are shown in Table 1 (all cases [ N = 366] and population controls) and Table 2 (symptomatic cases [ N = 261] and controls). Table 1 and Table 2 also show the odds ratios (OR), with 95% CIs for the two comparisons: all cases versus population controls and symptomatic cases (gynaecological pain, menstrual symptoms, or infertility) versus matched symptomatic controls. As expected, pain was more common in cases in both comparisons: OR 14.9, 95% CI = 10.1 to 21.9 versus population controls and OR 5.6, 95% CI = 3.9 to 8.1 versus symptomatic controls over 3 years’ data. Menstrual bleeding and timing symptoms were coded more commonly than in population controls, OR 3.8, 95% CI = 2.8 to 5.0 and 2.1, 95% CI = 1.4 to 3.2, but not in comparison with symptomatic controls, OR 1.0, 95% CI = 0.7 to 1.4 and 1.2, 95% CI = 0.7 to 1.9. Non-specific clinical features such as fatigue, vulvo-vaginal problems, and lower gastrointestinal symptoms were all more common in cases than population controls. Although simple tests such as full blood count were more common in cases than population controls, there was no significant difference in the symptomatic comparison. Genitourinary swab tests (presumably ordered because of the possibility that symptoms were due to pelvic inflammation) were more common in cases than controls in both comparisons. Occurrence of prescribed treatments In both the population and the symptomatic group comparisons, both analgesics (OR 3.0, 95% CI = 2.3 to 4.0 and OR 2.7, 95% CI = 1.9 to 3.9, in 3 years before index date Table 1. Numbers, proportions, and odds ratios (95% CI) for features in cases of endometriosis compared with population controls Occurrence of features in 3 years before index date a Occurrence of features at any time before index date a Cases ( N = 366) Controls ( N = 1453) Cases ( N = 366) Controls ( N = 1453) Specific features n % n % OR 95% CI n % n % OR 95% CI Subfertility 39 10.7 24 1.7 7.7 (4.4 to 13.3) 41 11.2 31 2.1 5.9 (3.6 to 9.7) Menstrual — bleeding 121 33.1 179 12.3 3.8 (2.8 to 5.0) 151 41.3 267 18.4 3.3 (2.6 to 4.3) Menstrual — timing 39 10.7 80 5.5 2.1 (1.4 to 3.2) 45 12.3 117 8.1 1.6 (1.1 to 2.3) Ovarian 24 6.6 7 0.5 13.7 (5.9 to 31.8) 25 6.8 11 0.8 9.8 (4.7 to 20.4) Pain 145 39.6 79 5.4 14.9 (10.1 to 21.9) 169 46.2 146 10.1 9.9 (7.1 to 13.6) Non-specific symptoms Fatigue 56 15.3 121 8.3 2.0 (1.4 to 2.8) 79 21.6 178 12.3 2.0 (1.5 to 2.7) Gynaecological 51 13.9 47 3.2 5.0 (3.3 to 7.7) 77 21.0 97 6.7 4.0 (2.8 to 5.6) Lower GI 104 28.4 144 9.9 3.7 (2.8 to 5.0) 126 34.4 213 14.7 3.3 (2.5 to 4.3) Upper GI 27 7.4 62 4.3 1.8 (1.1 to 3.0) 50 13.7 107 7.4 2.1 (1.4 to 3.0) Urinary 25 6.8 49 3.4 2.1 (1.3 to 3.5) 42 11.5 80 5.5 2.3 (1.5 to 3.5) Tests and procedures Full blood count 40 10.9 102 7.0 2.0 (1.2 to 3.2) 50 13.7 112 7.7 2.6 (1.6 to 4.2) Genital swabs 64 17.5 77 5.3 4.5 (3.0 to 6.7) 73 20.0 111 7.6 3.5 (2.5 to 5.0) Laparoscopy 42 11.5 13 0.9 14.6 (7.5 to 28.4) 47 12.8 15 1.0 13.9 (7.5 to 25.7) Thyroid function 53 14.5 112 7.7 2.4 (1.6 to 3.5) 67 18.3 132 9.1 2.8 (1.9 to 4.1) Ultrasound 14 3.8 5 0.3 12.3 (4.0 to 37.8) 14 3.8 11 0.8 5.0 (2.2 to 11.4) Treatments Contraception 201 54.9 716 49.3 1.3 (1.0 to 1.6) 234 63.9 800 55.1 1.5 (1.2 to 2.0) NSAID 171 46.7 276 19.0 4.8 (3.6 to 6.4) 191 52.2 393 27.1 3.8 (2.9 to 5.1) Analgesic 136 37.2 254 17.5 3.0 (2.3 to 4.0) 156 42.6 343 23.6 2.7 (2.1 to 3.5) SSRI 65 17.8 188 12.9 1.5 (1.1 to 2.0) 85 23.2 229 15.8 1.7 (1.2 to 2.2) Tricyclic 29 7.9 60 4.1 2.2 (1.3 to 3.6) 42 11.5 82 5.6 2.4 (1.6 to 3.6) a Index date: date of diagnosis for cases, date of diagnosis of matched case for controls. CI = confidence interval. Gynaecological = vulvo-vaginal symptoms, pelvic inflammation. Lower GI = pain, bloating, irritable bowel syndrome. NSAID = non-steroidal anti-inflammatory drug. OR = odds ratio. Ovarian = coded diagnosis of ovarian cysts and related conditions. SSRI = selective serotonin reuptake inhibitor and related antidepressants. Upper GI = dyspepsia, reflux, nausea. British Journal of General Practice, December 2017 e819 comparison) and NSAIDs (OR 4.8, 95% CI = 3.6 to 6.4 and OR 3.0, 95% CI = 2.1 to 4.2, in 3 years before index date comparison) were more commonly prescribed to cases than controls. When comparing cases and symptomatic controls, there was no association with antidepressant drugs (either tricyclic or SSRI and related). Composite features Table 3 shows the number and proportion of patients with at least one instance of each of the composite features over the 3 years before date of diagnosis/matching. Several composite features had high ORs when cases were compared with symptomatic controls: pain and menstrual symptoms within the same year (pain proximity menstrual [360]), OR 6.5, 95% CI = 3.9 to 10.6 and lower gastrointestinal symptoms occurring within 90 days of gynaecological pain (OR 6.1, 95% CI = 3.6 to 10.6). Episodes of gynaecological pain separated by at least 180 days were approximately eight times as likely in cases than symptomatic controls (OR 8.5, 95% CI = 4.3 to 16.9). Although pain or analgesic use on stopping contraception was suggested by some of the experts, these composite features occurred in less than 10% of cases, and with only moderate ORs of approximately 3. Occurrence of diagnostic features over the time prior to diagnosis Figure 1 shows plots of eight diagnostic features, describing the ORs for 3-year time windows with different intervals between the end of the 3-year window and the diagnosis/ matching date. Each plot compares cases with matched population controls and symptomatic cases with their matched symptomatic controls. In all plots, 95% CIs Table 2. Numbers, proportions, and odds ratios (95% CI) for features in cases of endometriosis compared with symptomatic controls Occurrence of features in 3 years before index date a Occurrence of features at any time before index date a Cases ( N = 261) Controls ( N = 610) Cases ( N = 261) Controls ( N = 610) Specific features N % N % OR 95% CI N % N % OR 95% CI Subfertility 39 16.1 52 8.5 2.4 (1.4 to 3.9) 41 16.9 64 10.5 1.9 (1.2 to 3.1) Menstrual — bleeding 121 49.8 304 49.8 1.0 (0.7 to 1.4) 151 62.1 443 72.6 0.7 (0.5 to 0.9) Menstrual — timing 30 12.4 64 10.5 1.2 (0.7 to 1.9) 34 14.0 111 18.2 0.7 (0.5 to 1.1) Ovarian 14 5.8 3 0.5 12.2 (3.5 to 42.7) 15 6.2 6 1.0 7.0 (2.7 to 18.1) Pain 145 59.7 148 24.3 5.6 (3.9 to 8.1) 169 69.6 241 39.5 4.0 (2.8 to 5.6) Non-specific symptoms Fatigue 45 18.5 84 13.8 1.4 (0.9 to 2.1) 66 27.2 138 22.6 1.3 (0.9 to 1.9) Gynaecological 41 16.9 34 5.6 4.2 (2.4 to 7.4) 64 26.3 68 11.2 3.6 (2.3 to 5.6) Lower GI 79 32.5 109 17.9 2.3 (1.6 to 3.2) 95 39.1 180 29.5 1.7 (1.2 to 2.3) Upper GI 24 9.9 51 8.4 1.3 (0.8 to 2.3) 44 18.1 87 14.3 1.5 (1.0 to 2.3) Urinary 20 8.2 29 4.8 1.8 (1.0 to 3.4) 36 14.8 64 10.5 1.5 (1.0 to 2.4) Tests and procedures Full blood count 34 14.0 82 13.4 1.2 (0.7 to 2.2) 42 17.3 97 15.9 1.4 (0.8 to 2.4) Genital swabs 43 17.7 71 11.6 2.2 (1.3 to 3.5) 50 20.6 90 14.8 1.9 (1.2 to 3.0) Laparoscopy 31 12.8 4 0.7 20.0 (7.0 to 57.1) 35 14.4 13 2.1 7.2 (3.7 to 14.1) Thyroid function 43 17.7 86 14.1 1.5 (0.9 to 2.4) 53 21.8 103 16.9 1.7 (1.1 to 2.7) Ultrasound 11 4.5 6 1.0 5.2 (1.6 to 17.0) 11 4.5 7 1.2 4.3 (1.4 to 13.0) Treatments Contraception 151 62.1 373 61.2 1.1 (0.8 to 1.5) 178 73.3 421 69.0 1.3 (0.9 to 1.9) NSAID 133 54.7 185 30.3 3.0 (2.1 to 4.2) 150 61.7 264 43.3 2.6 (1.8 to 3.7) Analgesic 100 41.2 142 23.3 2.7 (1.9 to 3.9) 116 47.7 203 33.3 2.3 (1.6 to 3.4) SSRI 43 17.7 115 18.9 1.0 (0.7 to 1.5) 57 23.5 148 24.3 1.1 (0.8 to 1.6) Tricyclic 20 8.2 37 6.1 1.5 (0.8 to 2.7) 29 11.9 58 9.5 1.3 (0.8 to 2.1) a Index date: date of diagnosis for cases, date of diagnosis of matched case for controls. CI = confidence interval. Gynaecological = vulvo-vaginal symptoms, pelvic inflammation. Lower GI = pain, bloating, irritable bowel syndrome. NSAID = non-steroidal anti-inflammatory drug. OR = odds ratio. Ovarian = coded diagnosis of ovarian cysts and related conditions. SSRI = selective serotonin reuptake inhibitor and related antidepressants. Upper GI = dyspepsia, reflux, nausea. e820 British Journal of General Practice, December 2017 are indicated. These show differing patterns. The plot for fertility problems (infertility) shows that until 1.5 years before diagnosis there was no association with a diagnosis of endometriosis, but from there the OR increased until about 0.5 years before diagnosis, at which point it stayed elevated. This is interpreted as indicating that the time delay from the occurrence of infertility to diagnosis is relatively short, presumably as infertility leads to referral including diagnostic laparoscopy. The plot for gynaecological pain shows that the OR was significantly elevated several years prior to diagnosis and that this increased in the year prior to diagnosis (at least in the population comparison). The two plots for non-specific symptoms (fatigue and lower gastrointestinal symptoms) show patterns of longstanding modest elevation. The bottom row of plots in Figure 1 shows two composite features: lower GI symptoms within 90 days of gynaecological pain and episodes of gynaecological pain >180 days apart. Although CIs for these composites were wider there was a suggestion of a trend over time in the lower GI plus pain combination.

Summary This study has two important new findings. First, the predictive value of several composite features for a subsequent diagnosis of endometriosis in routine records was evaluated. Second, for the first time, different time trends in the appearance of recorded clinical features of endometriosis were demonstrated. Strengths and limitations The choice of features as pointers used principles of feature selection based on expert input, 19 and methods of data consolidation and aggregation that have been developed for use with clinical data sources Table 3. Numbers, proportions, and odds ratios (95% CI) for composite features in the 3 years before diagnosis/matching a Comparison with population controls Comparison with symptomatic controls Composite feature Cases ( N = 366) Controls ( N = 1453) Cases ( N = 261) Controls ( N = 610) n % n % OR 95% CI n % n % OR 95% CI Pain during contraception 40 10.9 24 1.7 7.4 (4.3 to 12.7) 40 16.5 38 6.2 3.0 (1.9 to 5.0) Pain follow contraception (180) 17 4.6 8 0.6 8.5 (3.7 to 19.7) 17 7.0 17 2.8 3.1 (1.5 to 6.4) Pain exclusive contraception 105 28.7 55 3.8 14.2 (9.1 to 22.0) 105 43.2 110 18.0 4.3 (2.9 to 6.2) Menstrual during contraception 38 10.4 65 4.5 2.6 (1.7 to 4.1) 38 15.6 87 14.3 1.1 (0.7 to 1.8) Menstrual follow contraception (180) 14 3.8 8 0.6 7.0 (2.9 to 16.7) 14 5.8 17 2.8 2.0 (1.0 to 4.2) Analgesic during contraception 51 13.9 90 6.2 2.5 (1.7 to 3.7) 39 16.1 59 9.7 2.0 (1.3 to 3.1) Analgesic follow contraception (180) 27 7.4 26 1.8 4.5 (2.5 to 7.8) 21 8.6 21 3.4 2.8 (1.5 to 5.3) Analgesic exclusive contraception 116 31.7 68 4.7 12.0 (8.1 to 17.8) 116 47.7 132 21.6 3.9 (2.7 to 5.6) NSAID during contraception 56 15.3 92 6.3 2.9 (2.0 to 4.2) 48 19.8 68 11.2 2.0 (1.3 to 3.0) NSAID follow contraception (90) 27 7.4 28 1.9 4.0 (2.3 to 6.8) 21 8.6 19 3.1 3.0 (1.6 to 5.8) Pain proximity menstrual (360) 61 16.7 23 1.6 15.1 (8.5 to 26.6) 61 25.1 34 5.6 6.5 (3.9 to 10.6) Analgesic proximity menstrual (90) 29 7.9 19 1.3 6.3 (3.5 to 11.4) 29 11.9 30 4.9 2.6 (1.5 to 4.6) Analgesic proximity pain (90) 45 12.3 15 1.0 15.5 (8.0 to 30.1) 45 18.5 20 3.3 7.1 (4.0 to 12.5) NSAID proximity pain (90) 63 17.2 28 1.9 10.9 (6.7 to 17.7) 63 25.9 40 6.6 6.0 (3.7 to 9.7) Lower GI proximity pain (90) 48 13.1 12 0.8 15.9 (8.4 to 29.9) 48 19.8 24 3.9 6.1 (3.6 to 10.6) Lower GI proximity menstrual (90) 35 9.6 23 1.6 6.3 (3.7 to 10.7) 35 14.4 39 6.4 2.6 (1.6 to 4.1) Pain separated by > 180 days 36 9.8 14 1.0 12.5 (6.3 to 24.6) 36 14.8 14 2.3 8.5 (4.3 to 16.9) a Composite feature names follow the format X relationship Y [N] where relationship is defined as follows: X during Y; only used where Y = contraception. X = feature and occurs at least once after the onset date and before the expected offset date of at least one contraceptive prescription. X follow Y (N); N = number of days. Y = discrete time point event. X = feature and occurs between 1 and N days after Y. Where Y = contraception, N days relate to the expected offset date. X proximity Y (N); used where X and Y = discrete time point events and N is a number of days. X occurs between N days before and N days after Y. X exclusive Y; currently only used where Y = contraception: X = feature. X and Y are present but criteria for X during Y are never met. A single prescription of contraception occurring on the same day as a code for dysmenorrhoea would meet X exclusive Y criteria as X during Y requires X after the onset of contraception. X separated by >(N) days; two consecutive occurrences of X separated by more than N days. CI = confidence interval. GI = gastrointestinal. NSAID = non-steroidal anti-inflammatory drug. OR = odds ratio. Funding This study was funded by the Chief Scientist Office of NHS Scotland through its first health informatics call (reference HICG/1/25). The funder played no role in conducting the research or in writing the article. Ethical approval The study involved analysis of anonymised data. Access to the data was approved by the Research Applications and Data Management Team at the University of Aberdeen. Provenance Freely submitted; externally peer reviewed. Competing interests The authors have declared no competing interests. British Journal of General Practice, December 2017 e821 other than GP records. 17,20 This sequence of steps is broadly comparable with other recent approaches to the summarisation of clinical data. 20,21 An established anonymised GP record set was used that contained both diagnostic and symptom codes using the Read Code format, which means that the

is transferable to other research datasets and potentially to clinical use. There were limitations relating to the data, as the data were from standalone primary care records with no linkage to secondary care records, meaning that the reliability of GPs’ diagnoses of endometriosis could not be assessed. However, in the authors’ experience, GP practices tend not to code such diagnoses without specialist opinion. The data were more sparse than anticipated, with only around half of cases having cardinal clinical features of endometriosis recorded prior to diagnosis. This probably reflects the limited use of symptom codes by GPs, even in this database where a reason for consultation was meant to be given for each attendance. The rates of coding of procedures such as laparoscopy was surprisingly low; the authors suspect this is because GP practices had coded the findings of the laparoscopy rather than the procedure itself. Finally, as the duration of the database was shorter than a female’s reproductive period, a decision was made to exclude some females aged >35 years and diagnosed with endometriosis in order to maintain a focus on females for whom electronic health records were more likely to have data about earlier menstrual and related symptoms. Comparison with existing literature The authors are not aware of other studies that have looked for combinations of features in time as predictors of diagnoses in GP records. Although combinations of symptoms are commonly used in cancer prediction tools, these are usually simply recorded as present or absent, 22 whereas in this study temporal relationships were specified in order to increase the specificity of pointers. Other studies of endometriosis have only reported single items. 5 Implications for research and practice The composite predictors of a diagnosis of endometriosis variables reflect the patterns that clinicians observe, and, for the first time, they have been tested using data in routine GP records over time. These combinations — including pain and menstrual symptoms in the same year; pain and lower GI symptoms in the same 90 days; and episodes of pain separated by at least 6 months — are likely to be clinically useful, as pointers to a diagnosis in their own right. However, the fact that they can be derived from existing data means that they have potential to be included in diagnostic support software within GP records. 23 This study did not have sufficient cases to split the data into derivation and test sets, but future studies can use these composite features to test their predictive value in larger and better linked datasets. Additionally, machine learning techniques have a potential value in feature reduction and model selection. 24,25 Ultimately, the aim must be to apply these observations within predictive models for earlier referral and diagnosis of endometriosis. Infertility Gynaecological pain Years gap before index date 3.0 0.1 0.5 2.0 10.0 Odds ratio 2.5 2.0 1.5 1.0 0.5 0.0 Years gap before index date 3.0 0.1 0.5 2.0 10.0 Odds ratio 2.5 2.0 1.5 1.0 0.5 0.0 Fatigue Lower GI symptoms Years gap before index date 3.0 0.1 0.5 2.0 10.0 Odds ratio 2.5 2.0 1.5 1.0 0.5 0.0 Years gap before index date 3.0 0.1 0.5 2.0 10.0 Odds ratio 2.5 2.0 1.5 1.0 0.5 0.0 Analgesic prescription NSAID prescription Years gap before index date 3.0 0.1 0.5 2.0 10.0 Odds ratio 2.5 2.0 1.5 1.0 0.5 0.0 Years gap before index date 3.0 0.1 0.5 2.0 10.0 Odds ratio 2.5 2.0 1.5 1.0 0.5 0.0 Gynaecological pain and lower GI witin 90 days Episodes gynaecological pain >180 days apart Years gap before index date 3.0 0.1 0.5 2.0 10.0 Odds ratio 2.5 2.0 1.5 1.0 0.5 0.0 Years gap before index date 3.0 0.1 0.5 2.0 10.0 Odds ratio 2.5 2.0 1.5 1.0 0.5 0.0 OR versus population controls OR versus symptomatic controls OR versus population controls OR versus symptomatic controls OR versus population controls OR versus symptomatic controls OR versus population controls OR versus symptomatic controls OR versus population controls OR versus symptomatic controls OR versus population controls OR versus symptomatic controls OR versus population controls OR versus symptomatic controls OR versus population controls OR versus symptomatic controls Figure 1. Plots of OR for individual features over 3 years, by gap between the end of the 3-year window and the date of diagnosis/matching. Dotted lines indicate 95% CI for ORs. CI = confidence interval. OR = odds ratio.

The authors thank the expert clinicians and representatives of Endometriosis UK for their interviews. Discuss this article Contribute and read comments about this article: bjgp.org/letters e822 British Journal of General Practice, December 2017

1. Ballard K, Lowton K, Wright J. What’s the delay? A qualitative study of women’s experiences of reaching a diagnosis of endometriosis. Fertil Steril 2006; 86(5): 1296–1301. 2. Dunselman GA, Vermeulen N, Becker C, et al . ESHRE guideline: management of women with endometriosis. Hum Reprod 2014; 29(3): 400–412. 3. Pugsley Z, Ballard K. Management of endometriosis in general practice: the pathway to diagnosis. Br J Gen Pract 2007; 57(539): 470–476. 4. Staal AH, van der Zanden M, Nap AW. Diagnostic delay of endometriosis in the Netherlands. Gynecol Obstet Invest 2016; 81(4): 321–324. 5. Ballard KD, Seaman HE, de Vries CS, Wright JT. Can symptomatology help in the diagnosis of endometriosis? Findings from a national case-control study — Part 1. BJOG 2008; 115(11): 1382–1391. 6. Simoens S, Dunselman G, Dirksen C et al . The burden of endometriosis: costs and quality of life of women with endometriosis and treated in referral centres. Hum Reprod 2012; 27(5): 1292–1299. 7. Culley L, Law C, Hudson N, et al . The social and psychological impact of endometriosis on women’s lives: a critical narrative review. Hum Reprod Update 2013; 19(6): 625–639. 8. Abbas S, Ihle P, Köster I, Schubert I. Prevalence and incidence of diagnosed endometriosis and risk of endometriosis in patients with endometriosis- related symptoms: findings from a statutory health insurance-based cohort in Germany. Eur J Obstet Gynecol Reprod Biol 2012; 160(1): 79–83. 9. Lemaire GS. More than just menstrual cramps: symptoms and uncertainty among women with endometriosis. J Obstet Gynecol Neonatal Nurs 2004; 33(1): 71–79. 10. Nnoaham KE, Hummelshoj L, Kennedy SH, et al. Developing symptom-based predictive models of endometriosis as a clinical screening tool: results from a multicenter study. Fertil Steril 2012; 98(3): 692–701. 11. Gupta D, Hull ML, Fraser I, et al . Endometrial biomarkers for the non- invasive diagnosis of endometriosis. Cochrane Database Syst Rev 2016; (4): CD012165. 12. Nisenblat V, Bossuyt PM, Farquhar C, et al . Imaging modalities for the non- invasive diagnosis of endometriosis. Cochrane Database Syst Rev 2016; (2): CD009591. 13. Hirsch M, Begum MR, Paniz E, et al . Diagnosis and management of endometriosis: a systematic review of international and national guidelines. BJOG 2017; Jul 29. DOI: 10.1111/1471-0528.14838 14. Ballard K, Lane H, Hudelist G, et al . Can specific pain symptoms help in the diagnosis of endometriosis? A cohort study of women with chronic pelvic pain. Fertil Steril 2010; 94(1): 20–27. 15. Chapron C, Souza C, Borghese B, et al. Oral contraceptives and endometriosis: the past use of oral contraceptives for treating severe primary dysmenorrhea is associated with endometriosis, especially deep infiltrating endometriosis. Hum Reprod 2011; 26(8): 2028–2035. 16. Sleeman D, Moss L, Aiken A, et al . Detecting and resolving inconsistencies between domain experts’ different perspectives on (classification) tasks. Artif Intell Med 2012; 55(2): 71–86. 17. Reis BY, Kohane IS, Mandl KD. Longitudinal histories as predictors of future diagnoses of domestic abuse: modelling study. BMJ 2009; 339: b3677. 18. Burton C, Cochran AJ, Cameron IM. Restarting antidepressant treatment following early discontinuation — a primary care database study. Fam Pract 2015; 32(5): 520–524. 19. Sleeman D, Moss L, Sim M, Kinsella J. Predicting adverse events: detecting myocardial damage in intensive care unit (ICU) patients. KCAP 2011, the Sixth International Conference on Knowledge Capture. Banff, Alberta, Canada: 2011. New York: ACM Press: 73–79. DOI: 10.1145/1999676.1999690. 20. Feblowitz JC, Wright A, Singh H, et al . Summarization of clinical information: a conceptual model. J Biomed Inform 2011; 44(4): 688–699. 21. Hirsch JS, Tanenbaum JS, Lipsky Gorman S, et al . HARVEST, a longitudinal patient record summarizer . J Am Med Inform Assoc 2015; 22(2): 263–274. 22. Hamilton W. The CAPER studies: five case-control studies aimed at identifying and quantifying the risk of cancer in symptomatic primary care patients. Br J Cancer 2009; 101(Suppl 2): 80–86. 23. Nurek M, Kostopoulou O, Delaney BC, Esmail A. Reducing diagnostic errors in primary care. A systematic meta-review of computerized diagnostic decision support systems by the LINNEAUS collaboration on patient safety in primary care. Eur J Gen Pract 2015; 21(Suppl): 8–13. 24. Mitchell TM. Machine learning . Boston: WBC/McGraw-Hill, 1997. 25. Darcy AM, Louie AK, Roberts LW. Machine learning and the profession of medicine. JAMA 2016; 315(6): 551–552. British Journal of General Practice, December 2017 e823