Among incident cancer cases, we have mapped the presenting symptom signatures of 38 cancer sites and described the cancer site case-mix of 83 presenting symptoms. Certain presenting symptoms are typically concentrated in cancers of the same body system or region, and vice versa. When the symptom signature of a given cancer is dominated by a single symptom, the cancer case-mix of that symptom is also dominated by the same cancer; conversely, relationships between presenting symptom and cancer site are much weaker for cancers with broader symptom signature. The cancer site case-mix of certain symptoms (e.g. musculoskeletal symptoms) varies by sex.

Evidence on the presenting symptom signature of 15 cancer sites has been reviewed previously [10]. Generally, relevant prior evidence is concentrated on single cancer sites; in contrast we have examined the symptom signature of 38 cancers simultaneously. Acknowledging this difference, our findings concord with prior evidence, although they expand to an additional 23 cancers sites with little or prior population-based evidence on their presenting symptoms (such as laryngeal, liver, melanoma, mesothelioma, oral, penile, sarcoma, small intestinal, testicular, thyroid, vaginal and vulval) [10].

Consistent with previous literature, we have found breast and bladder cancers to have narrow symptom signatures [10], and haematological [11], pancreatic [12], and renal cancers [10] to have broad symptom signatures. Haemoptysis has a relatively high predictive value for lung cancer, but it only occurs in 20% of patients with lung cancer [13]; consistent with this prior evidence, we found that a fifth of lung cancer patients in our study population presented with haemoptysis. As described previously, we have found that chronic lymphocytic leukaemia patients typically had no recorded symptoms, which concords with prior evidence indicating that this cancer is often detected asymptomatically [11]. Prostate, renal, liver and thyroid cancers also had high percentages of diagnosis without recorded symptoms, consistent with prior knowledge about detection via opportunistic screening or incidental identification in many patients [14]. The observation that ocular and oral cancers also have a high proportion of non-recorded symptoms is novel, and could indicate detection of those cancers outside primary care (e.g. opticians and dentists). In keeping with prior evidence, most cancers arising after upper or lower abdominal symptom presentations related to cancers of the abdomen (e.g. upper and lower GI, hepato-pancreato-biliary, urological, prostate and other male organs/gynaecological) though around one in five related to non-abdominal sites [15, 16]. Concurring with prior evidence, we have observed a diverse symptom signature for colon and rectal cancers [17].

In brief, the findings concord with prior evidence but amplify it substantially in respect of number of presenting symptoms examined and the range of cancer sites considered.

We covered a wide range of symptoms and cancer sites in a population-based incident cohort of patients with cancer. However, there are several limitations to consider. This was a case-only analysis (only patients with diagnosis of cancer were included). While this is an inherent feature of all epidemiological studies using cancer registry data, it is important that this is borne in mind for interpretation. For example, no inferences can be made about the predictive value of certain symptoms for specific cancers.

By the design of the NCDA questionnaire, GPs were asked to record the first presenting symptom(s) that prompted the suspicion of cancer. Since the surveys were filled by the GPs retrospectively, it is possible that certain symptoms more closely related to the patient’s diagnosis were recorded in the audit. It is also possible that other symptoms, particularly non-specific ones, were present and did relate to the underlying cancer but not deemed to do so by the GP, and therefore were not recorded. However, GPs had access to both coded and free-text data in the patients’ records—a unique feature of NCDA study, which mitigates concerns about reliance on structured (coded) fields in primary care electronic health record data sources [18]. The symptoms recorded relate to the symptoms presented to the GP which may differ from the symptoms experienced at symptom onset.

Although we present associations between individual symptoms and individual cancer sites, we have also grouped symptoms and cancer sites to provide higher-level summaries. The definition of these groups is normative and chiefly guided by anatomical considerations; by its nature includes a degree of heterogeneity.

Three main translational implications arise from the findings.

Considering research implications, the results provide foundational evidence that can be used to validate the completeness of phenotyping of cancer symptoms in electronic health record sources, or profile associations between presenting symptoms and diagnostic process measures, such as investigation use.

Considering implications for public health or clinical practice, the findings can guide decision-making about the choice of target symptoms in symptom awareness campaigns, and inform their evaluation, regarding the range of cancer sites where changes in diagnostic pathways and intervals may be observable. Similarly, they can guide investigation strategies in patients presenting with specific symptoms, for example, prioritising certain tests over others (e.g. endoscopy over imaging), given differences in the expected probability of specific cancer sites. Further, novel/emerging diagnostic technologies, such as multi-cancer early detection tests, could be preferentially deployed on symptoms potentially associated with a wider range of cancer sites. As an example, Multi-Cancer Early Detection (MCED) tests provide information on whether a cancer signal was detected (yes/no) and up to two predicted ‘cancer site origins’, i.e. suspected site of underlying cancer [19]. Clinicians may be able to complement such information with evidence on the associations between presenting symptoms and the likely distribution of cancer sites in cases presenting with that symptom, to further inform investigation strategies and test sequencing. Additionally, our study could motivate future studies into examining the probability of specific cancer sites in cancer patients conditional on their presenting symptoms, which could further improve the diagnostic accuracy and usefulness of information that can be derived by MCED tests.

Considering implications for public policy, the findings emphasise the importance of considering the overall risk of cancer (across body organs and systems) in symptom-based referral guidelines for suspected cancer. Instead, current UK guideline recommendations (issued in 2015) chiefly relate to symptoms of specific cancer sites [6], meaning that, as we show, the broader cancer site case-mix of different symptoms (particularly vague symptoms) is not appreciated. There is no equivalent single body of guidelines for presenting symptoms in the US setting, although evidence indicates that diagnostic delays in symptomatic patients subsequently diagnosed with cancer are comparable to those seen in Europe, and that the predictive values of different symptoms among presenters are also similar [20, 21]. Recent evidence demonstrates that the predictive values of three common vague symptoms, i.e. weight loss, fatigue and abdominal pain, do not exceed the 3% normative referral threshold used by NICE when individual cancer sites are considered on their own, but do so when all cancer sites are considered together [22,23,24].