Data

The English National Cancer Registry (CR) was individually linked to Clinical Practice Research Datalink (CPRD). These data were released under national statutory approvals from The Confidentiality Advisory Group (CAG): PIAG 1-05(c)2007, PIAG 3-06(f) 2008 and national ethical approvals from the Research Ethics Committee (REC): 13-LO-0610, 08-H1102-46.

CPRD is a complex, comprehensive database of anonymised electronic health records for individual patients registered at general practices (GP) across the UK. It contains patient demographic information, clinical diagnoses, prescriptions, immunisations, test results and certain information on health and health-seeking behaviours. It is one of the largest sources of continuous patient records available and is broadly representative of patient and practice characteristics in the UK. Research-quality (up-to-standard, UTS) data are available for 13 million patients, five million of whom are currently registered, submitted to the central database by over 600 UK primary care centres. Linkage of CPRD primary care data with other patient level datasets, including cancer registrations, is available for English practices within CPRD who have consented to participate in the linkage scheme.

Case definition and control selection

Cases were defined as patients aged 15–99 years diagnosed with primary pancreatic cancer (ICD-10 code C25) between January 1, 2005 and June 30, 2009 whilst registered with a CPRD practice who had been submitting UTS data for at least two years before their diagnosis (S1 Fig). Controls were patients diagnosed with an unrelated primary cancer at least 18 months after the index date (date of diagnosis of matched PC case) who were similarly registered within a CPRD practice submitting UTS data during the identical chronological period. We excluded patients diagnosed with cancers of the lip, oral cavity and pharynx (ICD-10 codes C00-14), digestive organs (C15-26), respiratory and intrathoracic organs (C30-39), breast (C50) and, female genital organs (C51-58) which share at least one symptom or risk factor with pancreatic cancer. Each case was matched to four controls on the basis of sex and birth cohort ± 2.5 years.

Variable coding

We derived symptoms and health statuses of patients known to be associated, or possibly associated with a subsequent diagnosis of pancreatic cancer during the two-year period prior to diagnosis (S1 Table). First, an exhaustive list of medical or product (drug) codes associated with each symptom was derived from the Code Browser using broad search terms. Code lists for frequently studied symptoms and drug use were also obtained via correspondence with authors of published papers as well as from the LSHTM Data Compass [20], a curated digital repository of research outputs produced by staff and students at the London School of Hygiene & Tropical Medicine (LSHTM) and their collaborators. We also searched for clinical codes on CALIBER [21] which is a comprehensive, open access resource for UK electronic health records data.

Medical and drug prescription codes along with their event dates up to two years prior to index date of pancreatic cancer were extracted using Clinical and Therapy files provided by CPRD. We also searched for useful information in the Additional, Referral and Test files. Codes were assigned to an appropriate symptom to build an analysable data structure. These symptoms aimed to be as comprehensive as possible; derived from both extensive literature searches as well as discussions with our clinical collaborator (SPP). After establishing a clean and exhaustive code list for each symptom and drugs prescribed, the codes were checked for non-repetition across all the different variables. Fifty-three symptoms were coded in this fashion.

We derived patients’ status in relation to smoking, body mass index (BMI) [22], diabetes and, alcohol consumption using previously-published code lists. We allocated ‘ever smoker’ status if patients had been recorded as being a current smoker or ex-smoker in the past two years prior to index date, whilst ‘ever heavy drinker’ status was similarly attributed to patients recorded as current or ex heavy alcohol consumers. Any recording of BMI in the previous 24 months (or height and weight on the same day) was extracted. If the most recent BMI prior to cancer diagnosis was greater than or equal to 30, the patient’s status was marked as obese. We allocated diabetic status to a patient if (s)he recorded any of the diabetes associated medical codes or were prescribed anti-diabetic drugs during the time window of interest. In addition to medical codes on weight loss entered by the GP, patients were also coded as having experienced weight loss as a symptom if his/her most recent BMI was either five or more per cent lower than their highest recorded BMI or the first BMI in the two years before cancer diagnosis. We used ecologically-derived measures of income deprivation for each person using quintiles of the income domain score from the English Indices of Multiple Deprivation 2004, 2007 and, 2010 [23]. Deprivation categories were derived from the score temporally closest to each person’s index date on the basis of their residential address. The consultation file was used to calculate the total number of appointments a patient had with a doctor at the GP surgery within the time period under study.

All medical records referring to administration and routine examination codes were excluded since they did not contain any clinical information. Symptoms relating to the eyes, ears, skin and, skeleton were also excluded. Finally, we visually examined all the remaining non-assigned codes occurring more than ten times in the database. These included symptoms of cough, sore throat, dizziness, Polymyalgia rheumatica, gout, venesection and, Salicylate prophylaxis. In discussion with our clinical collaborator (SPP), we excluded most of these codes along with the codes which had fewer than ten clinical events within the whole cohort during the entire study period.

Ethics

This pilot study was conducted on pre-existing linked data in order to evaluate the feasibility of this approach. These data were released under national statutory approvals from The Confidentiality Advisory Group (CAG): PIAG 1-05(c)2007, PIAG 3-06(f) 2008 and national ethical approvals from the Research Ethics Committee (REC): 13-LO-0610, 08-H1102-46. The study protocol was approved by the London School of Hygiene & Tropical Medicine Ethics Committee on April 25, 2019 (Ref 17053).

Statistical analyses

We fitted models in the machine learning setting using R software [24]. Our aim was to derive a risk score for developing PC for each patient at time points prior to diagnosis. The outcome was the probability (between zero and one) that a patient would develop the disease, derived as a function of the input variables which included all symptoms, status variables (obesity, smoking, alcohol and, diabetes) age, sex, social deprivation and, the number of consultations in the time window of interest.

Initially, all input variables for the full 24 months before diagnosis were included. The model was then refitted for each monthly interval between 1 and 20 prior to diagnosis, disregarding each time information collected after the assessment time point. For example, the model fitted for month one excluded any information recorded less than one month before diagnosis, but included information between months 2–24; the model for 20 months excluded all information recorded prior to 20 months, but included information for months 21–24. Multivariate logistic regression (LR) and random forest (RF) models were fitted within the supervised machine learning setting. We regarded the four patient statuses as non-time varying by recording them as binary (yes/no) variables if they occurred within the time window of interest. Symptoms were fitted as continuous variables by including the count of the number of times a symptom or drug was reported during the time window of interest.

The models were trained on 75% of the data (the training set) and then assessed on the remaining 25% (the test set) by the derivation of the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. Cases and their matched controls were kept together in one or other of the sets. The model’s performance measures (AUC, sensitivity and specificity) were assessed at different threshold values (cut-off points) between 0.1 to 0.5. If the predicted probability of a patient’s outcome was higher than the threshold value, then they were considered ‘high-risk’ of PC. The analyses were conducted by age in order to compare younger patients (aged 60 years and under) to older patients (more than 60 years).

Descriptive analyses

We identified 1,144 patients with primary pancreatic tumours eligible for inclusion. Four age- and sex- matched controls were identified for 1,139 cases, whose male-to-female ratio was 0.957 and whose median age at diagnosis was 71 years. We were unable to find four controls for the remaining five cases: these were females with mean age of 93 years all of whom died within seven months from diagnosis. These cases were excluded.

We derived the percentage of cases and controls who reported each symptom at any time during the 24 months prior to pancreatic cancer diagnosis (cases) or the index date (controls) (Table 1). Pearson’s chi-square test [25] was used to test if these observed differences between cases and controls for each symptom were by chance. Abdominal pain, irritable bowel syndrome and, constipation symptoms were reported around four times more frequently among cases than controls while gastrointestinal conditions and diabetes were twice as common. There was a negligible difference in the prevalence of genitourinary and immunological disorders. The prevalence of each symptom in every monthly interval before diagnosis among cases and controls varied by symptom. For example, reporting of jaundice was very different for cases and controls but only very close to diagnosis (S2–S9 Figs). By contrast, higher prevalence of gastrointestinal problems amongst cases was evident from 24 months onwards.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Percentage of cases and controls reporting symptoms or statuses within the 24-month period prior to diagnosis.

https://doi.org/10.1371/journal.pone.0251876.t001

Month by month, cases reported more symptoms than controls throughout the whole period of observation (S10 Fig). The cumulative median number of symptoms in each time window was higher among cases but still substantial among controls (S10 Fig). The number of symptoms rose towards the diagnosis in both groups but especially in cases. Amongst patients reporting at least one symptom in a given month (on average, 80% of cases and 70% of controls) cases reported much higher total number of symptoms than controls in the same period (S10 Fig).

Multivariable regression

Multivariate logistic regression (LR) and random forest (RF) models presented similar results, indicating that the choice of model was not important. For simplicity we henceforth report outputs from only logistic regression models.

The performance of the models was checked in the 25% test set in every time window up to 20–24 months preceding diagnosis. An optimal cut-off for the predicted probability of outcome, where AUC is maximal, was determined separately for models in each time window. The AUC, sensitivity and, specificity were plotted corresponding to these optimal cut-offs against varying time periods. The AUC (which represents a measure or degree of separability of a true positive from a false positive at various threshold values) was optimal (highest) when using all available 24 months, including the period immediately prior to diagnosis, and decreased as the time window of analysis moved further away from diagnosis (Fig 2A). Sensitivity (the ability to correctly detect a patient who developed pancreatic cancer) was maximized at 73% at the time of diagnosis, dropping to 35% in younger and 53% in older patients in time windows prior to 6 months from diagnosis (Fig 2B). Specificity (the ability to correctly detect a patient who did not develop PC) ranged from 75%-88% across all time windows for all patients (Fig 2C). The model’s performance varied by age, with higher sensitivity and lower specificity in earlier months in the age-specific analyses, whilst the AUC was similar.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Model parameters obtained in each time window (0–24,1–24,…,20–24 months) prior to diagnosis.

a) AUC (%) b) Sensitivity (%) c) Specificity (%).

https://doi.org/10.1371/journal.pone.0251876.g002

Since we were principally interested in maximizing sensitivity early in the course of the disease, further analyses were conducted on age-specific models for the time window more than 12 months prior to diagnosis which displayed the maximal AUC: 20 months for younger patients and 17 months for older patients (S2–S5 Tables). For each of these time windows we plotted age-specific model’s performance measures against various values in order to establish the optimal cut-off point. For patients aged 60 or under, this was 0.3 (Fig 3) where the AUC was 65.6%, the sensitivity of this model was 72.5% and, the specificity was 58.7%. In the population of patients over 60 years, the optimal cut-off was 0.21 where the AUC was 60.9%, the sensitivity was 65.1% and, specificity of 56.8% (Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. AUC, sensitivity and specificity of the models with % of the population recommended for biomarker (Stage 2) testing.

https://doi.org/10.1371/journal.pone.0251876.g003

Our algorithm identified 41.3% (Fig 3) of younger patients and 43.2% (Fig 3) patients being classified into a ‘high-risk sub-population’ (Fig 1). Using the most recently published incidence for England [7,8] this equates to 10,923 persons per pancreatic cancer present in younger patients and 740 in older patients. Were a biomarker with a sensitivity of 87.5% [17] applied, around 30 older and 400 younger persons would be identified as ‘potential pancreatic cancer patients’ (those referred on for imaging and potentially endoscopy) leading to the early detection of around 63% of pancreatic cancers in younger patients and 57% in older patients. (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Parameters of the optimal logistic regression models envisioned in the context of the hypothetical multi-stage screening model*: England 2017.

https://doi.org/10.1371/journal.pone.0251876.t002

Diabetes was the principal influential risk factor in younger patients. (S2 and S4 Tables). Since we learned the importance of diabetes in both age groups, we observed 89% of the younger and 96% of the older diabetic patients correctly predicted to develop PC, as opposed to 68% of the younger and 47% of the older non-diabetics.

Public health context

It is widely accepted that pancreatic cancer is not a good candidate for mass population-based screening due to its low incidence in the general population. As small as 10% of the patients diagnosed with pancreatic ductal adenocarcinoma (PDAC) have either a strong family history of PC, hereditary pancreatitis or a particular genetic syndrome [1]. The remaining 90% of PDAC patients are non-inherited cases who overwhelmingly display no alarm symptoms until the tumour is well-advanced.

Strengths and limitations

This study has a number of strengths. First, we used prospectively collected data and a high-quality database. Under-recording or inaccurate reporting may be present but is unlikely to be differential. We evaluated temporal patterns by establishing the predictive accuracy of our model at separate time points throughout the two-year period leading up to diagnosis. We used the machine learning setting to carry out statistical analyses, rather than multivariable regression methods or conditional logistic regression as has previously been implemented [26,27]. Machine learning techniques are more powerful in settings such as this one where they are more likely to identify numerous weak signals which are only predictive when used in complex combination with each other. Both of the prediction models (LR and RF) produced similar outcomes suggesting that our results are valid irrespective of the model choice. This is an advantage of applying machine learning techniques in our study setting which confirms comparable predictions using two different regression models.

The one major limitation of our study is the poor specificity of our models arising principally from the use of cancer patients as controls. Our controls were selected from cancer patients diagnosed later in time with unrelated malignancies. Despite the fact that we ensured that the controls’ own diagnosis was at least 18 months after that of their matched case, they nevertheless reported increasing numbers of symptoms of interest over the study period in a fairly similar fashion to that of the cases. This implies that the controls are not representative of the general population, and explains why our models predict a high number of false positives. However, in the light of this limitation, our results are still promising: the models achieve a relatively high sensitivity and the specificities of the final models are in excess of 55%. Repeating the analyses using full population-based controls would lead to a much higher specificity and is the first priority in progressing this research.

Our results suggest that further age differentiation is required to truly evaluate the utility of this approach in patients under 60 years, which would be possible in a larger dataset. In particular, it would be important to evaluate performance separately for patients aged 40–49 and 50–59 amongst whom the potential effectiveness of our models is currently obscured by the very low incidence amongst 15-59-year olds. Initial calculations show that if the model parameters for ages 15–59 held for the limited age group of 40–59 years (median age 55 years) 1,290 persons per pancreatic cancer would have been identified as ‘high-risk’ and 47 as ‘potential cancer patients’ (compared to 10,923 and 394 for the age group 15–59).

Further limitations are relatively minor. Alcohol consumption and smoking habits are almost certainly under-reported in primary care [28,29], but arguably accurately reflect what is known to the GP at any given time and thus are as good as clinical knowledge of any other symptom. Patient status was recorded as ‘diabetic’ if this disease was reported in the time window of interest which did not permit us to distinguish between long-term or recent onset disease.

Comparison with existing literature

Previous authors have highlighted putative features associated with PC diagnosis such as jaundice, abdominal pain, back pain, gastric problems, weight loss, malaise and new-onset diabetes [26,30]. Whilst our results are consistent with these findings, we have examined whether it is possible to predict future PC diagnosis based on the presence or absence of symptoms or abnormalities present more than 12 months before diagnosis, ignoring late stage symptoms. Comparison with existing literature is, therefore, limited.

Implications for future research

This study is preliminary, and thus needs to be expanded and extended in order for these findings to be confirmed. First, it is essential that future analyses use population-based controls so that true specificity of the algorithm within a population-based database can be derived. Second, analyses should differentiate between diabetics and non-diabetics, as well as in a more refined manner by age. A notable result was the relative importance of diabetes, over time-varying symptoms, in predicting later pancreatic cancer diagnosis, which is consistent with previous research [31]. Since studies have established a particularly increased risk amongst new onset diabetics [1], stratified analyses by diabetic history hold promise to improve the predictive ability of the model for all patients. Third, testing the algorithm in ‘real time’, where a primary care database is interrogated as if this algorithm were in use is important to understand its cost-effectiveness. In reality, the greatest potential of this algorithm lies within a multiple-testing model, where a number of different biomarkers for several different malignancies are evaluated simultaneously within the same patient. This study is an important first step to establishing such a programme.

Conclusion

We have demonstrated that it is possible to discriminate a high-risk sub-population of patients whom are more likely to go on to be diagnosed with pancreatic cancer from routine primary care records. Our results could lead to increasing early diagnosis through a multi-stage screening model which utilises recently developed biomarkers applied to the ‘high-risk’ population identified using this approach. Further work is certainly required to confirm, refine and evaluate the potential use of these findings in practice.