Cut Ovarian Cancer Detection Costs with Risk-Based Triage

The study focused on assessing pathways to detect invasive OC, excluding borderline tumours (see Supplementary Table S1 for ICD-10 codes), in primary care.

Pathways for OC diagnosis in primary care

We considered six pathways for OC diagnosis among women presenting in primary care with symptoms of possible OC. Firstly, we examined the NICE recommended pathway (Pathway 1) of CA125 test, followed by pelvic USS if CA125 ≥ 35 U/mL (Fig. 1a).

Secondly, two further CA125-USS sequential pathways using Ovatools-informed [10] risk-based triage or equivalent age-adjusted CA125 thresholds were studied (Fig. 1b). In Pathway 2, patients first undergo a CA125 test and Ovatools risk is calculated. If the Ovatools invasive OC probability is <1% patients are not investigated further for OC; if it is between 1% and 2.9% (moderate risk) they are followed up with pelvic USS; and if ≥3% (high risk) they are directly referred on an urgent suspected OC pathway for further assessment. The 1% cancer risk threshold has been used by NICE when selecting symptomatic patients for primary care investigations (such as chest X-ray in patients with symptoms of possible lung cancer) and the 3% threshold aligns with the level recommended for urgent suspected cancer referrals in symptomatic patients [7]. Pathway 3 mimics Pathway 2 but instead uses age-adjusted CA125 thresholds with OC detection accuracy matching the corresponding Ovatools probabilities of ~1% and ~3% (see Supplementary Method 1 for details of the age group specific thresholds and their accuracy) [10].

Thirdly, three further CA125-USS concurrent pathways were studied, where patients undergo both CA125 testing and a pelvic USS in primary care (Fig. 1c). In Pathway 4, patients are referred if either the Ovatools invasive OC probability is ≥3% or OC is suspected on USS. Pathway 5 uses age-adjusted CA125 threshold matching Ovatools probability of 3% and Pathway 6 uses CA125 ≥ 35 U/mL threshold in place of the Ovatools probability.

This study included women with records indicating they underwent a CA125 test and/or a pelvic/transvaginal USS between April 1st, 2013, and December 31st, 2017, in Clinical Practice Research Datalink (CPRD) Aurum. The index date for the study was the earliest valid CA125 test or USS record within the study period. CPRD Aurum contains routinely-collected data from 40 million patients from 1500 UK general practices [12], linked at the patient level to Hospital Episode Statistics (HES), Diagnostic Imaging Data (DID), death registration data and National Cancer Registration and Analysis Service (NCRAS) data [13]. Women with a previous USS or CA125 test within 365 days before the index date or with a record for OC (including borderline ovarian tumour) at any time before the index dates were excluded. Women with a USS record but neither a CA125 record nor ovarian related symptoms were also excluded as the indication for USS is not always recorded within CPRD. Patients were followed until March 31st, 2021 [14]. As in previous studies, we assumed all patients with cancer at the time of CA125 testing or USS were diagnosed within 1 year [15].

The base-case cost-effectiveness analysis used a narrower target population of CA125 tested women, while certain model parameters were estimated using the broader study population.

The decision analytic model

The decision analytic model consists of a decision tree to represent the diagnostic pathways and costs in primary care, and a Markov model to capture long-term survival, quality of life (QoL) and inpatient care costs. The structure aligns with the model that informed NICE recommendations [11]. The decision tree estimates patients’ probabilities of the following outcomes after assessment in primary care, namely, (1) invasive OC detected and referred to secondary care (true positive), (2) invasive OC undetected and patient reassured (false negative), (3) no invasive OC and reassured (true negative), and (4) no invasive OC but referred to secondary care (false positive). To parametrise the decision tree, we used the accuracy data for CA125 (with a 35 U/mL threshold) and Ovatools, stratified by age ( < 50 years and ≥50 years), as the accuracy changes significantly with age (Table 1) [10]. We applied the same USS accuracy metrics as used in the study informing NICE and the economic evaluation of OC pathways in UK secondary care [11, 16]. Disease incidence and stage were estimated based on patient characteristics from the study population using logistic regression models (Supplementary Method 2).

We estimated the effect of higher cancer detection rates in primary care by modelling the cancer stage shift [17]. We applied it to patients who were diagnosed at the late stage of cancer if they were in the false negative group in the current pathway but in the true positive group in the new pathway. We assumed that a fraction of additionally detected ‘previously late stage’ cancers would ‘shift’ to an early stage at diagnosis in the new pathway. This fraction was estimated using the relative risk ratio of late-stage diagnosis incidence between screen detected cases and clinically detected cases among symptomatic women in the UKCTOCS study [18] (Supplementary Method 3).

The Markov model includes five states: three entry states—no cancer, early-stage cancer, and late-stage cancer – and two absorbing states – death caused by cancer (cancer death) and death not caused by cancer (non-cancer death) (Supplementary Fig. S1). The model uses annual cycles. The transition probabilities from the cancer states to cancer death were estimated using flexible parametric survival models, which modelled time from cancer diagnosis to cancer death using data from the study population. As a high proportion of patients with raised CA125 have other cancers [8], the survival models were estimated separately for OC and cancers including lower gastrointestinal (GI) cancer, uterine cancer, lung cancer and pancreatic cancer, and other cancer types were modelled as a whole. Age and cancer stage are key covariates. Missing cancer stage data was imputed (Supplementary Method 4). No detailed cancer progression was modelled, but the models of cancer death risk subsume cancer progressions. The estimated survival models were used to predict cancer deaths in cancer patients in the first eight years after diagnosis in the Markov model. The national statistics of mortality rates by age, sex and cause (cancer and non-cancer) were used to predict separately all non-cancer deaths and cancer deaths beyond eight years for women diagnosed with cancer, and all non-cancer deaths and cancer deaths for women not diagnosed with cancer. The models showed good performance in validation on internal and external data (Supplementary Method 5).

Three categories of costs, in 2022 British pounds (£), were included in the analyses. Firstly, we considered the costs for services and diagnostic tests in primary care pathways. These include the costs for the general practitioner (GP) time, nurse time, the CA125 test, and the USS (see Table 1 for cost details and sources) [19,20,21]. We assumed that each patient received an initial face-to-face GP consultation, followed by a CA125 test (and USS in concurrent pathways). In sequential pathways, if patients’ CA125 test results indicated an USS, they received a GP telephone follow-up to inform them of the result and arrange an USS appointment, and if the USS was abnormal, they received a further GP consultation before being referred to secondary care. In cases where cancer existed but was not detected in the initial diagnostic pathway, we assumed the diagnostic process in primary care was repeated once.

Secondly, we estimated the long-term hospital inpatient care costs from hospital records, including cancer treatment, using the Hospital Episode Statistics (HES) Health Resource Group (HRG) data. HRG codes were mapped to NHS reference costs inflated to 2022 prices [22]. Following methods used previously [23, 24], separate Two-Part Generalised Linear Models were fitted for patients with a certain cancer type as those in the survival models, and patients without cancer diagnosed. Duration since the first diagnosis was defined as a categorical variable with annual levels from the year of diagnosis to four or more years after diagnosis. Cancer stage at diagnosis, current age, ethnicity and socioeconomic deprivation were included as covariates. These cost models were integrated into the Markov model for prediction of inpatient care costs since diagnosis. See Supplementary Method 6 for further details about cost estimation.

Finally, we added hospital costs for referrals without an eventual cancer diagnosis, because the above inpatient care cost models do not effectively capture hospital cost for those referred to hospital but not ultimately diagnosed with cancer. We separately added costs for patients who did and did not undergo surgery for benign diseases. We used data from the ROCkeTS study to determine the surgery rate among patients referred from primary care who were not diagnosed with any cancer (Table 1) [25], and applied the activity-weighted cost of “Treatment with interventions for Non-Malignant Gynaecological Disorders” in the NHS reference costs [22]. For non-surgery cases, we assumed all patients received an outpatient consultation, a CA125 test and a USS.

We estimated QoL associated with patient characteristics and disease histories using a linear regression model of EuroQoL-5 Dimension (EQ-5D) utility using data from UK Biobank [26]. The EQ-5D utility generally ranges from around −0.5 for the worst health state to 1 for full health where 0 is equivalent to death [27]. Histories of different cancers were defined as categorical variables with levels of no such cancer, cancer diagnosed within one year, and cancer diagnosed more than one year previously. The QoL model was integrated into the Markov model for prediction of QoL over lifetime and estimation of quality adjusted life years (QALYs). See Supplementary Method 7 for further details. As UK Biobank does not provide data on cancer stage, we applied an additional QoL decrement of 0.046 derived from the literature on top of the gynaecological cancer-related coefficients in the QoL model to reflect the impact of late-stage cancer [28]. For patients referred to secondary care without a cancer diagnosis, if they received surgery, we assumed a reduction of 0.04 in QoL in the year of surgery [29], and an improvement of 0.008 in annual QoL thereafter [30] (Table 1).

Cost-effectiveness analysis

Base-case analysis focused on women with CA125 records to inform the real-world cost effectiveness of implementing Ovatools thresholds in place of the current CA125 cut-off, while we also conducted analysis on the wider population including also symptomatic women with USS but without CA125 testing. We conducted cost-effectiveness analyses from the perspective of the UK health services. We used the decision model to simulate diagnostic outcomes, remaining life years (LYs), QALYs and costs for the study populations over a lifetime (or until 110 years of age) assuming they followed a particular primary care pathway. We applied a 3.5% annual discount rate to costs and QALYs.

We used two approaches to compare the results of different pathways. First, we compared the NICE recommended Pathway 1 with each alternative pathway to derive QALYs gained and additional costs and identified dominated alternatives. Second, we calculated incremental cost-effectiveness ratios (ICERs) based on comparisons of moving to increasingly effective and costly alternative pathways, excluding the dominated alternatives [31]. We did not directly compare the Ovatools pathways with their corresponding age-adjusted threshold pathways, as the differences between them were minor due to the age-group specific thresholds being aligned with Ovatools’ risk-based accuracy.

Given that the sensitivity of Ovatools at the 1% risk threshold is lower than that of CA125 at the 35 U/mL threshold for individuals aged < 50 years but higher for those aged ≥ 50, we analysed the data separately for these two age groups.

Sensitivity and scenario analyses