Study population and case identification

The study population included all dogs under first opinion care at participating VetCompass clinics in 2019. Dogs were defined as under veterinary care if they had at least one electronic health record (EHR) —such as a clinical note, treatment, or bodyweight measurement—recorded during 2019 or in both 2018 and 2020. Available data fields include species, breed, date of birth, sex, and neuter status, clinic ID, corporate veterinary group, as well as date-specific information on free-text clinical notes, bodyweight, and billed item treatments.

Potential haemangiosarcoma cases (candidate cases) were identified using the following search terms in the clinical notes: (haemangios*, hemangios*, HSA, angiosa*, haemangiosarcoma~1, hemangiosarcoma~1, hameangios*). A haemangiosarcoma case was defined as a dog with evidence of a final clinical diagnosis (with or without histopathological confirmation) of haemangiosarcoma in the EHRs. As not all cases underwent a definitive histopathological diagnosis, the diagnosis of haemangiosarcoma was considered a clinical diagnosis in some dogs. Only incident cases (those diagnosed between January 1 2019 and December 31 2019) were retained for this study. Ethical approval for the VetCompass programme was granted by the Social Science Research Ethical Review Board at the Royal Veterinary College (SR2018−1652).

Clinical record coding

A total of 4,997 candidate cases were randomly ordered and each EHR was manually reviewed to determine if it matched the haemangiosarcoma case definition. Candidate cases were excluded if haemangiosarcoma was listed only as a differential diagnosis or if there was no final confirmed diagnosis. Haemangiosarcoma cases had additional clinical information extracted and coded. Information from the coded variables was then grouped into further categories for case description, survival time calculations and regression modelling (Table 2). Fixed values of breed, sex, neuter status, date of birth and clinic and corporate group IDs were downloaded automatically from the EHR. The information coded and its groupings are shown in Table 2. Cases were restricted to those first diagnosed at ≥5 years old, as haemangiosarcomas (amongst other cancers) have been shown to be rare in young dogs [3,24].

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Information extracted from clinical records of haemangiosarcoma cases diagnosed in 2019 in VetCompass EHRs. The ‘Variable’ column shows what information was captured from EHRs and what it was transformed or grouped into for regression analyses (‘Grouped variables’).

https://doi.org/10.1371/journal.pone.0316066.t002

Due to the aggressive and invasive nature of visceral haemangiosarcoma it was considered unlikely to be able to establish whether a tumour location was primary or metastatic mass based on the EHR information so tumours were additionally grouped more broadly into whether there was any tumour present at the spleen, heart, liver or cutaneous tissues (‘any tumour location’ variables in tumour specifics row in Table 2). A complete list of information coded can be found in S1 Table Supplementary Mateiral.

Survival time

Survival was calculated from date of first diagnosis to death or until the end of the study or the animals last available EHR (if the last EHR was before December 31, 2023). Animals that had not died were censored for survival analysis on December 31, 2023. Median survival time (MST, days) and 1-year survival rates (range, % and 95% confidence intervals, CI) were calculated separately for: all haemangiosarcoma cases and all cases that lived ≥1 day after first diagnosis. These calculations were also performed separately by tumour location for splenic, hepatic, cardiac and cutaneous tumours. Kaplan–Meier survival (i.e., time-to-event) curves were constructed to describe time to death from haemangiosarcoma diagnosis for all cases and then separately for each tumour location group.

Factors associated with dying on day of diagnosis

Binary logistic regression modelling was used to ascertain whether there were distinct associations between a range of factors and an outcome variable of either dying on the day of presentation and or dying ≥1 day after presentation.

Initial univariable binary logistic regression modelling assessed for potentially significantly associated variables. Explanatory variables with liberal associations with survival beyond day of diagnosis (p < 0.2) were carried forward for multivariable binary logistic regression modelling [27]. A correlation matrix was built and Variance Inflation Factor (VIF) was used to assess for collinearity between potentially significant explanatory variables. Substantial collinearity was considered present with correlation ≥0.7, VIF > 5 [27]. For collinear variables, each of the variables were assessed individually in the model and the variable with smallest model AIC value was retained for further modelling. The multivariable model was built using a manual stepwise forwards construction, approach where variables with the smallest p-value from the likelihood ratio test (LRT) and/or smallest Akaike Information Criteria (AIC) value were included first. Variables were retained in the model if the model was deemed improved with variable inclusion through a significant likelihood ratio test (LRT) result. AIC values were used to compare non-nested models that were built with different collinear variables to determine which model balanced goodness of fit and complexity best. Pairwise interactions were assessed for variables in the final model and retained if significant (LRT p < 0.05). Performance of the final model was assessed through Area Under the Curve (AUC) and model fit to data was assessed by a Hosmer-Lemeshow test, where p > 0.05 indicates that the model adequately fits the data [27]. A random effect of clinic ID was added to the model. Intraclass coefficient (ICC) was calculated to compare level of variance within a clinic to level of variance between clinics. Statistical significance was set at the 5% level [27].

Cox regression modelling

Two separate survival analyses were undertaken using Cox proportional hazards regression modelling. In the first model, all dogs with ≥1 day survival after diagnosis (n = 407) were considered to evaluate the hazard of death from both clinically diagnosed and histopathologically diagnosed haemangiosarcoma cases. In order to explore whether factors associated with hazard of death for dogs clinically diagnosed with haemangiosarcoma were consistent in lab confirmed haemangiosarcoma cases, a second model was built using only dogs with a histopathological diagnosis of haemangiosarcoma (n = 272). Explanatory variables with liberal univariable association with hazard of death with haemangiosarcoma (p < 0.2) were carried forward for multivariable Cox proportional hazards regression modelling [27]. A correlation matrix for potentially significant explanatory variables was assessed, as described for the binary logistic regression model approach. Clinical signs were excluded from multivariable Cox regression modelling due to their lack of independence. Both age at diagnosis (years) and maximum tumour size (cm) recorded were both normally distributed and appeared to increase linearly with hazard of death and both variables were explored as a continuous variable and as categorical variables by quartiles. The categorical maximum tumour size variable was taken forward into multivariable Cox regression modelling due to the large number of missing values of tumour size (n = 258) whereas age was retained as a continuous variable as there were no missing records for this.

Multivariable model building used a manual stepwise forwards construction approach. Variables were retained in the model if the model was deemed improved with variable inclusion through a significant likelihood ratio test (LRT) result and the smallest AIC value for separate models built with collinear variables. Pairwise interactions between final model explanatory variables were evaluated and retained if their inclusion significantly improved the model (LRT p value <0.05). The proportional hazards assumption (i.e., the assumption that the hazard ratios (HRs) for category variables remain constant over time) was tested through assessment of Schoenfeld residuals and visualisation of survival curves in Kaplan-Meier plots for each variable [27]. Schoenfeld residuals were assessed for individual variables in the final model and for the global model. If the global model did not violate the proportional hazards assumption (p > 0.05) then this model was used. Model performance and fit to the dataset was assessed through Concordance Index and Xu and O’Quigley R² values [27–29]. Statistical significance was set at the 5% level [27].

All data cleaning and analyses was performed in R Studio (R Core Team, Vienna, Austria) using the following packages: survminer, dynpred, ggsurvfit, ggplot2, finalfit and coxr2.

Study population

From the study population of 2.2 million dogs, there were 4,997 candidate haemangiosarcoma cases. Of these, 788 cases met the case definition for newly diagnosed haemangiosarcoma in 2019. Once age restricted to dogs over five years, the median age at first diagnosis was 10.6 years (interquartile range, IQR, 9.1–12.1 years). Nearly three-quarters of the population were neutered (n = 577,73.4%). Male dogs comprised 53.5% of the cases (n = 422) and females 45.6% (n = 359). The three most common breeds were: Labrador Retriever (n = 96, 12.2%), German Shepherd Dog (n = 84, 10.7%) and Staffordshire Bull Terrier (n = 43, 5.5%). Overall, crossbred dogs were the most common dog type (n = 177, 22.5%).

There were 29 tumour locations recorded with the most common ones being spleen (n = 489, 62.1%) and liver (n = 198, 25.1%) followed by cutaneous haemangiosarcoma (n = 146,18.5%).

Most EHRs reported at least one clinical sign at presentation, with 28 (3.5%) dogs having no clinical signs recorded at presentation. A wide range of clinical signs were recorded, with 64 distinct clinical signs noted. Most commonly, clinical signs were non-specific: lethargy (n = 309, 38.6%), anorexia/hyporexia (n = 287,35.8%) and abdominal distension (n = 197, 25.0%) (Fig 1).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Grouping of clinical variables coded from electronic health records of dogs presenting with haemangiosarcoma to VetCompass clinics in 2019 (n = 788).

Bar chart colours indicate the group that an individual variable was grouped into: Clinical signs grouped according to their main body system affected (top bar plot), diagnostic tests performed (middle bar plot) and treatment options undertaken (bottom bar plot).

https://doi.org/10.1371/journal.pone.0316066.g001

The most commonly performed diagnostic tests were abdominal ultrasound (n = 547, 68.3%) and blood tests (n = 485, 60.5%). Effusion analysis was performed in 166 patients (20.7%), of these the majority were abdominocentesis (n = 143, 17.8%) and less commonly pericardial effusions (n = 23, 2.9%). In one-third of patients, a confirmatory histopathological diagnosis was performed (n = 277, 34.6%). None of the age-excluded (dogs <5 years old) had histopathology performed.

For staging purposes, 30.2% (n = 238) of patients had their tumour size recorded, only 21.4% (n = 169) of patients had presence of metastases recorded and of these dogs, only 7 (0.9%) recorded metastases to lymph nodes. Therefore, tumour staging was not taken further in the survival analyses. Median tumour size was 6.0 cm (IQR = 3.5–10 cm, range 0.2–23.0 cm).

Surgical excision was the most commonly performed treatment (n = 276/788, 35.0%), with splenectomy being the most common type of surgery (n = 172/276, 62.3%). Palliative care (analgesia or steroid administration) was the next most common treatment (n = 102, 12.9%). Chemotherapy (either high- dose or metronomic) was uncommon (n = 42, 5.3%). Of those, anthracycline-based (doxorubicin or epirubicin) protocols were most commonly used (27 dogs, 64.3% of chemotherapy dogs) followed by metronomic cyclophosphamide (16 dogs, 38.1% of chemotherapy patients).

Whether death was the result of euthanasia or unassisted was recorded in most patients (n = 676, 85.8%), with euthanasia being the mechanism of death in most cases (n = 624, 92.3%).

Survival time

Survival time across all clinically diagnosed haemangiosarcoma cases ranged from 0–1789 days (IQR 0–74 days). The median survival time (MST) of all cases was 9 days from diagnosis (95% CI 5–15 days), with a 1-year survival rate of 12% (9.7–15%) (Table 3). Dogs with cardiac and hepatic haemangiosarcoma had the lowest MST at 0 days (95%CI 0–0) for both (Fig 2). The 1-year survival rate for these locations was 3% (95% CI 0.8–12%) and 3.9% (95% CI 1.8–84%) respectively. The MST for splenic haemangiosarcoma was 4 days (95% CI 0–9 days). Cutaneous haemangiosarcoma had the longest MST at 119 days (95% CI 85–248 days), with a 1-year survival rate of 35% (95% CI 28–44%).

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Summary table of case numbers, median survival time (MST), interquartile range (IQR) and 1-year survival rate for clinically diagnosed haemangiosarcoma cases diagnosed in first opinion practices in VetCompass in 2019. Cases are grouped by location: any, cardiac, splenic, hepatic and cutaneous. Cases are divided further to separate all cases and cases with ≥1 day of survival time post-diagnosis.

https://doi.org/10.1371/journal.pone.0316066.t003

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Kaplan-Meier survival curve estimating the survival time of clinically diagnosed haemangiosarcoma cases diagnosed in first-opinion practices in VetCompass in 2019 separated by location affected: cardiac, cutaneous, hepatic, splenic.

Day 0 designates day of first diagnosis. Shaded areas indicate the 95% CI for each curve and crosses indicate censored cases.

https://doi.org/10.1371/journal.pone.0316066.g002

Of those dogs with a confirmed histopathological diagnosis of haemangiosarcoma (n = 272), MST was 105 days (84–133 days), with a 1-year survival rate of of 28% (22–34%). For hepatic, splenic and cutaneous haemangiosarcoma locations, the MST was longer for those with a clinical diagnosis (hepatic MST: 79 days (95% CI 35–112 days), splenic MST: 80 days (95% CI 64–106 days), cutaneous MST: 248 days (95% CI 119–478 days)). The 1-year survival rate for these locations was: 7.2% (95% CI1.9–26.7%), 16% (95%CI 10.1–24.2%) and 43.2% (95% CI 34.8–53.6). The MST and 1-year survival rate for histopathologically confirmed cardiac haemangiosarcoma could not be calculated as there was only one case.

Factors associated with death on day of diagnosis

Due to the high proportion of dogs not surviving for ≥1 day after diagnosis (n = 381/788, 48.3%) identified in survival time calculations it was hypothesised that there were two distinct populations of haemangiosarcoma cases. Therefore, binary logistic regression was performed to ascertain risk factor differences between dogs with <1 day of survival (n = 381) and those with ≥1 day of survival after diagnosis (n = 407). In univariable logistic regression comparing these two groups, 37 variables were found to be potentially significant (LRT p- value <0.2) that were taken forward to multiple logistic regression (S2 Table in Supplementary Material). Of these variables, there were 52 pairs of correlated variables which were assessed in non-nested models to decide on which of the pair to include. The final model comparing differences between the dogs with <1 day survival and ≥1 day survival retained 8 variables: if the patient had surgery or medical treatment, haemangiosarcoma location, whether they visited a referral centre, whether there were mass associated clinical signs present), whether there were haematological abnormality clinical signs present, whether abdominal diagnostic tests were performed and whether there were any clinical signs recorded at presentation. Additionally, corporate group was forced into the model to account for the structure of the data and Clinic ID was retained as a fixed effect (Fig 3). The model showed no evidence of poor fit to the data (Hosmer-Lemeshow p > 0.05) and good discrimination (AUC = 0.93). There was a small amount of clustering at the clinic level (ICC = 0.11).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Forest plot of final multivariable logistic regression model of haemangiosarcoma cases diagnosed in first-opinion practices in VetCompass in 2019 who died on day of diagnosis (‘Day 0’, n = 381) and those who died ≥1 day of survival post diagnosis (‘Day ≥1’, n = 407).

Day 0 and Day ≥ 1 columns represent number of animals and percentages for each category in the model. OR indicates the odds ratio for each category and its 95% confidence interval (CI) and p-value.

https://doi.org/10.1371/journal.pone.0316066.g003

Dogs that received any surgical or medical management of their haemangiosarcoma were more than 25 times more likely to die on day of diagnosis than those that did not (OR=26.93, 95% CI 17.22–43.41). The presence of haematologically abnormal clinical signs at presentation was associated with double the likelihood of dying on day of diagnosis (OR=2.29, 95% CI 1.43–3.76). Dogs with no clinical signs recorded at presentation were more than ten times less likely to die on day of diagnosis (OR=0.09, 95% CI 0.01–0.44). Dogs with mass associated clinical signs at presentation were half less likely to die on day of diagnosis (OR=0.52, 95% CI 0.31–0.88).

Dogs who visited a referral centre as part of their treatment were 10 times less likely to die on day of diagnosis (OR=0.10, 95% CI 0.04–0.23). A cutaneous location was associated with a seven-fold reduction in odds of dying on day of diagnosis compared to a splenic haemangiosarcoma (OR=0.14, 95%CI 0.03–0.23). Dogs with any abdominal diagnostic tests performed were half as likely to die on day of diagnosis (OR=0.53, 95%CI 0.28–1.00). None of the corporate group variables were significant.

These large variations in odds ratios between dogs that died on day of diagnosis and dogs that survived for ≥1 day indicated that these two dog groups were distinct and not suitable to be combined for survival analyses. In order to evaluate factors associated with survival with haemangiosarcoma diagnosis beyond first day of diagnosis, only dogs with ≥1 day of survival time were taken forwards.

Hazard for death with clinically diagnosed haemangiosarcoma in dogs that survive for ≥1 day after diagnosis

Of the 63 variables examined in univariable analyses, 41 were liberally significant (LRT p-value <0.2). Variables describing clinical signs at presentation were not taken forward due to their lack of independence as factors, leaving 43 factors to be assessed in multivariable modelling (S3 Table in Supplementary material). Of these, 28 variables were potentially correlated which were assessed in non-nested models for inclusion.

In the final Cox proportional hazards regression model, five variables were retained: whether a patient had any form of surgery, location of haemangiosarcoma, sex, maximum tumour size and the corporate group the clinic belonged to. Corporate group was not significant but was forced into the model to better account for the structure of the data. Additionally, a frailty term of clinic ID was incorporated to further account for data structure. The overall model and individual independent variables were assessed for violation of proportional hazards. All individual variables and the overall model satisfied this assumption. The final model performed well, with a pseudo-R² of 0.26 and a Concordance Index of 67.

Hazard of death increased as the maximum tumour size increased, with greatest hazard in dogs with haemangiosarcomas between 8−23 cm when compared to those with 0.2–4.3 cm tumours (HR = 1.94 95% CI 1.19–3.18). Male dogs had an increased hazard of death (HR = 1.32, 95%CI 1.05–1.65) compared to female dogs. Dogs with any surgery performed had a reduction in hazard of death when compared to those that didn’t (HR = 0.48, 95%CI 0.38–0.62). When compared to cardiac haemangiosarcoma cases, dogs with a haemangiosarcoma at other locations had reduced hazard of death with haemangiosarcoma, although this was only significant for cutaneous. Dogs with a cutaneous haemangiosarcoma had a decrease in hazard of death versus dogs with cardiac haemangiosarcoma (HR = 0.53, 95% CI 0.40–0.69) (Fig 4).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Forest plot of final mixed effects multivariable Cox proportional hazards regression results for factors associated with time to death with haemangiosarcoma in cases diagnosed in first-opinion practices in VetCompass in 2019 with ≥1 day of survival post diagnosis (n = 407).

Cases (%) column indicates the number of cases for each category in the model. HR indicates the hazard ratio for each category and its 95% confidence interval (CI) and p-value. The final model included a frailty term of Clinic ID.

https://doi.org/10.1371/journal.pone.0316066.g004

Hazard for death with histopathologically confirmed haemangiosarcoma in dogs

Of the 62 variables examined in univariable analyses, 38 were liberally significant (LRT p-value <0.2). Variables describing clinical signs at presentation were not taken forward due to their lack of independence as factors, leaving 29 factors to be assessed in multivariable modelling (S4 Table in Supplementary material). Of these, 15 variables were potentially correlated which were assessed in non-nested models for inclusion.

In the final Cox proportional hazards regression model, seven variables were retained: whether a patient had any form of surgery, location of haemangiosarcoma, maximum tumour size, whether a patient visited a referral centre, whether soft tissue metastases were identified, dog genotype grouping and the corporate group the clinic belonged to (Fig 5). Corporate group was not significant but was forced into the model to better account for the structure of the data. Additionally, a frailty term of clinic ID was incorporated to further account for data structure. The overall model and individual independent variables were assessed for violation of proportional hazards. All individual variables and the overall model satisfied this assumption. The final model performed well, with a pseudo-R² of 0.39 and a Concordance Index of 70.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Forest plot of final mixed effects multivariable Cox proportional hazards regression results for factors associated with time to death with haemangiosarcoma in cases histopathologically diagnosed in first-opinion practices in VetCompass in 2019 with ≥1 day of survival post diagnosis (n = 272).

Cases (%) column indicates the number of cases for each category in the model. HR indicates the hazard ratio for each category and its 95% confidence interval (CI) and p-value. The final model included a frailty term of Clinic ID.

https://doi.org/10.1371/journal.pone.0316066.g005

Surgery, haemangiosarcoma location and size were consistent between this model of histopathologically diagnosed haemangiosarcoma cases and all haemangiosarcoma cases. Dogs that had surgery performed had a reduced hazard of death (HR = 0.39,95%CI 0.25–0.61), when compared to those that did not. Dogs with cutaneous haemangiosarcoma had decreased hazard of death compared to splenic haemangiosarcoma cases (HR = 0.60,95%CI 0.43–0.83). Larger tumour size was associated with an increased hazard of death, this hazard was greatest for tumours between 8−23 cm when compared to 0.2–4.3 cm tumours (HR = 2.59, 95%CI 1.42–4.71). Visiting a referral centre was associated with a two-fold reduction in hazard of death (HR = 0.46, 95% CI 0.31–0.69). Dogs of the genotype group 2 (primarily consisting of spaniel, hound and poodle breeds) had decreased hazard of death (HR = 0.52, 95% CI 0.31–0.89) when compared to crossbred dogs (Fig 5).

Survival time after diagnosis with haemangiosarcoma

Visceral haemangiosarcomas are commonly regarded as aggressive neoplasms with a poor prognosis [1,4]. Short median survival times (MST) were consistently seen across all haemangiosarcoma locations in this study. When compared to existing studies, each haemangiosarcoma location had a substantially shorter MST in this study. Whilst dogs with cutaneous haemangiosarcoma had the longest survival time of all locations here (MST = 119 days) it was still shorter than previous studies, which reported MSTs between 172–1189 days [6,13,14,23]. Likewise, in the literature cardiac and splenic cases still have very short MST (7–189 days and 30–259 days, respectively) but these were substantially longer than what was seen here (0 days and 4 days, respectively) [7,9,10,11,30].

In most previous studies exploring haemangiosarcoma survival, studies included small case numbers primarily treated in specialist centres and presence of referral bias is likely [6–9]. This could have affected survival time through the availability of more advanced therapies, owners more motivated to pursue intensive treatment options and less financial concerns if attending a referral centre. Additionally, patients attending a referral centre might have had to present to their primary veterinarian clinically stable enough for referral. Furthermore, some studies excluded patients that did not meet a minimum survival length or did not survive to discharge which would have skewed for a longer MST [7,15]. When MSTs were calculated for cases with ≥1 day of survival after diagnosis or cases with a histologically confirmed diagnosis in this Vet Compass population, MSTs were similar to those reported in previous studies [7,9,10,30,31].

The potential reasons for short survival times are multifactorial. Haemangiosarcomas are reported to be generally aggressive tumours, with high metastatic rates [4]. In the study population here, a low proportion of dogs were receiving chemotherapy and many presented with severe clinical signs leading to potential inferences about low quality of life in haemangiosarcoma patients. However, the common prevailing view of extremely poor prognosis for haemangiosarcoma could be promoting frequent euthanasia at presentation and therefore leading to a self-fulfilling prophecy and low survival times. From reading clinical records it is difficult to be certain on the nuances of why euthanasia has been performed, with factors such as perceived prognosis, financial considerations and emotional distress being more difficult to capture than patient clinical presentation. As a result, it is not clear the proportional contribution of effects from these non-clinical factors on a decision to euthanise are challenging to derive and therefore the impact on survival times reported here.

For cutaneous haemangiosarcoma, the MST reported in the current results was at the lower end of previously reported MSTs in the literature [6,13,14,23,31]. This may have reflected that our categorisation of cutaneous haemangiosarcomas encompassed a wider range of non-visceral, more aggressive forms such as subcutaneous and intramuscular haemangiosarcomas. Additionally, solar radiation induced forms of haemangiosarcoma have been associated with increased survival time than non-solar induced [1,18]. Whilst the solar induced aspect of cutaneous haemangiosarcoma biology was not assessed for in cases here, due to reduced sunlight hours, intensity and UV levels, solar-induced haemangiosarcoma cases may be less frequent in northern Europe, meaning survival time may be reduced in cutaneous haemangiosarcomas in this current UK study [32,33].

In previous referral centre studies tumour staging has been shown to be a key prognostic factor [1]. However, tumour staging in accordance with adapted WHO guidance (tumour size, evidence of regional and distant metastases) was rarely recorded in clinical records in this study. Therefore, it was not possible to explore how the MSTs reported here relate to cancer stage.

Factors associated with dying on day of diagnosis

The binary logistic regression model aimed to explore clinico-pathologic and treatment factors associated with dogs that died on day of diagnosis versus dogs that survived for at least one day. Two factors were associated with dying on day of diagnosis: having surgical or medical management of their haemangiosarcoma and the presence of haematologically abnormal clinical signs. The management of haemangiosarcoma being associated with increased odds of death on day of diagnosis seems counter-intuitive initially, however this most likely reflects unstable patients requiring urgent surgery (for example, an actively bleeding splenic mass requiring emergency surgery) being associated with poorer outcomes. Likewise, the reduced likelihood of visiting a specialist centre in dogs that die on day of diagnosis also supports this, as dogs may need to be clinically stable before travelling for specialist care. As anticipated, based on survival times in existing literature [1,6,14], presence of a cutaneous haemangiosarcoma was associated with reduced odds of dying on day of diagnosis. Perhaps expectedly, an absence of clinical signs at diagnosis was associated with increased odds of surviving longer than one day. It seemed plausible that these would be more likely to be cutaneous lesions as these would be less likely to have clinical manifestations, however most of these cases were actually splenic (n = 22/28). Although the case numbers for this variable are small, this might be of clinical relevance to practitioners finding an incidental splenic mass on a work-up.

This model indicated that dogs dying on day of diagnosis had different associations with both modifiable and unmodifiable haemangiosarcoma factors than those that survived longer.

Factors associated with time to death with haemangiosarcoma

Previous survival analyses on haemangiosarcoma tended to focus on single locations for haemangiosarcoma. However, factors identified in this current study represent associations across all types of haemangiosarcoma. Cutaneous haemangiosarcomas were at significantly reduced hazard of death when compared to splenic haemangiosarcoma cases. The association of cutaneous haemangiosarcoma location with lowest hazard of death was consistent with median survival times reported in the literature across haemangiosarcoma locations [1,6–8,15], but this has not been explored through multi-location haemangiosarcoma survival analyses with a single study before.

In the current study, increasing tumour size was associated with increased hazard of death. Whilst larger masses (>6 cm) have been shown to be associated with a reduced survival time in cutaneous haemangiosarcoma [6], in several previous splenic haemangiosarcoma studies size has not been associated with prognosis [7,15]. Reasons for this lack of association in previous studies was survival to discharge post-splenectomy as an inclusion criterion might select for smaller masses as these might make better surgical candidates. Tumour size has not consistently been found to be a predictor of splenic malignancy as there is overlap between benign (specially hematomas) and malignant lesions, as well as between malignant lesions [26]. Tumour size may only be associated with prognosis for certain haemangiosarcoma locations. Since other studies exploring tumour size have focused on a single haemangiosarcoma location, our model which includes all locations makes interpreting this size association more complicated. For example, cardiac haemangiosarcomas have a very short survival time but are limited by cardiac anatomy to the extent they can grow to [9]. Both location and tumour size associations were consistent for both the model built with all clinically diagnosed cases and the model of only histologically diagnosed cases.

Male dogs had an increased hazard of death with haemangiosarcoma in the clinically diagnosed cases model. An association between sex and haemangiosarcoma survival time or diagnosis does not appear to be frequently identified in haemangiosarcoma studies [9,16,34]. Besides cancers of reproductive organs with their obvious sex associations there does not appear to be strong evidence for sex being associated with canine cancer survival time. This contrasts with humans, where men are often reported at increased risk of diagnosis and also poorer survival times across several cancers [35–37]. In human oncology, the exact mechanism of sex on both diagnosis and survival is not well understood [38]. There are sex hormone associations but also lifestyle and environmental differences between human sexes which contribute to this variation [35–37]. Sex hormones are involved in the regulation of physiological process such as stem cell renewal and the immune response and androgen receptor expression has been identified in angiosarcomas (a human neoplasm similar to haemangiosarcoma in dogs), indicating a potential sex hormone association for this cancer [38]. However, the extent to which the sex difference is translatable to veterinary is unclear given how commonplace neutering is, which minimises the potential role of sex hormones [39]. Neuter status was not retained in the final hazards model, further indicating that there are other factors beyond sex hormones contributing to this finding.

No significant effect on survival time was seen in the current study when medical treatments such as chemotherapy or alternative medicines (such as Yunnan baiyao or Coriolus versicolor) were utilised. The use of these therapeutics was uncommon (40 dogs for chemotherapy, 9 for alternative medicine), likely leading to a lack of power to detect an effect. However, as there is currently minimal evidence to suggest alternative medicines improve survival time it is also probable this association would not exist in the data here [22,34].

Previous studies reported some breed associations with haemangiosarcoma diagnosis but there is limited work evaluating the breed associations on survival after diagnosis [17,40]. Only cutaneous haemangiosarcomas has been previously reported to have some breed associations with survival time [14]. When histologically confirmed cases were examined, rather than variation in hazard for single dog breeds, dogs of the genotype group that consisted of spaniel, poodle, dachshund and beagle breeds were found to have a reduced hazard of death [26]. These were not the same breeds identified to have increased survival by Nóbrega and colleagues (2019) [14]. No breed associations were identified in the model of all haemangiosarcoma cases, nor with the survival time overall, in agreement with other existing literature [6,19].

Due to the very short survival times and high uptake of euthanasia in this study, it must be considered that these factors are not the primary drivers of survival time.

Limitations

Whilst working with large volumes of clinical records in VetCompass has the advantage of providing high number of cases there are some limitations to working with anonymised, retrospective records. For example, full laboratory records such as histopathology on masses or blood tests are often not available in the EPRs. In addition, the current first opinion study included many patients that did not have a definitive diagnosis pursued (only one-third had histopathology performed) or blood samples taken (only 60% of cases). This meant that the study was unable to assess the effect of variables associated with prognosis identified in other studies such as the mitotic count of a mass, Ki67 expression or thrombocytopaenia [8,12].

Relying on clinical diagnoses rather than exclusively histopathological diagnosis could have led to inclusion of misdiagnosed patients, as not all bleeding spleens are malignant masses and not all malignancies are haemangiosarcoma. Application of the ‘double two-thirds’ rule (where two-thirds of splenic masses are malignant and of these a further two-thirds are haemangiosarcoma) here indicates that the potential for misdiagnosis due to a lack of definitive diagnosis could have had an impact on survival time and factors associated with it identified here [11]. Meta-analyses by Schick and Grimes (2022) indicated that both splenic malignancies and haemangiosarcoma might be more common than the rule suggests [41] (Schick &Grimes, 2022). Achieving a definitive diagnosis can prognosticate better for owners and avoid a potentially unwarranted euthanasia (if benign). However, as this is often not feasible in acute, critical clinical presentation scenarios and in first opinion practice, presumptive diagnoses were included in the current analyses. Given the variation in proportions of biopsies between haemangiosarcoma locations in this study, three-quarters of cutaneous haemangiosarcoma cases were biopsied here versus just 2% and 28% of cardiac and splenic cases, it seems plausible that the severity of clinical presentation and the ability of first-opinion vets to perform haemangiosarcoma surgeries would have affected the uptake of biopsies. [1] Additionally, cytological diagnosis has limitations (such as blood contamination of samples) and an invasive biopsy may not be considered justifiable in patients with advanced disease [1]. Furthermore, histopathology alone can be insufficient to differentiate splenic haemangiosarcoma from other sarcomas, requiring immunohistochemistry confirmation which has additional time and cost implications [1]. Thus, if the study only included histologically diagnosed dogs, a large proportion of dogs would be excluded and survival would be biased to dogs with early stage haemangiosarcoma and longer survival times.

Euthanasia will undoubtedly have impacted survival risk factors. However, this study was interested in reporting the survival pattern overall and therefore both dogs euthanised due to the disease and dogs that died from the disease need to be included.

The reported effect of surgical removal on haemangiosarcomas on patient survival time varies widely. Whilst surgical removal has been shown to be potentially curative for cutaneous masses [6], it has been shown to lead to increased survival time in other locations with varying magnitude [1,7,9,19]. Therefore, the effect of surgery on survival time was a variable of interest for the current study. Unfortunately, due to the first opinion nature of these clinics and the fact that many haemangiosarcoma surgeries are highly complex and potentially require specialist surgeons (e.g., liver lobectomy) the effect of specific surgical procedures on survival time could not be assessed. As surgical procedures other than splenectomy were performed infrequently in the current caseload, all surgeries were all grouped into a binary variable which led to a loss of resolution between individual surgical types. Insurance status likely has an effect on both diagnosis and survival time. Unfortunately, these data are often not up-to-date in EHRs. Due to this, the relationship between insurance status and haemangiosarcoma cases was not explored.

Whilst these findings are likely applicable across wider UK first-opinion veterinary practices, it is unclear how applicable these findings are to other countries in the world. Factors such as breed genetic differences in other countries, access to diagnostics, standards of care and owner attitudes to euthanasia and treatment might vary between countries and lead to different associations than those seen here.

Although the mechanism of death was recorded in most EHRs in this study, it was still not possible to disentangle the multiple potential reasons for euthanasia such as perceptions around poor prognosis, financial considerations or whether the perceived acute quality of life deterioration prompted euthanasia. From the clinical notes in this study, it was not possible to determine whether dogs that died on the day of presentation and had surgery (n = 29) were euthanised intra-operatively or after exploratory surgery, only that the majority were recorded to have ‘inoperable tumour/s’ (n = 26) and the majority of these were euthanised (n = 27) rather than died (n = 2). Clinical notes pertaining to euthanasia consults are often brief and not able to capture non-clinical aspects. Therefore, qualitative research might be useful to explore haemangiosarcoma euthanasia considerations. Additionally, a small prospective study by Stewart and colleagues (2020) found that dogs presenting with splenic mass rupture (both benign and malignant) had more favourable outcomes than existing literature informing views on prognosis would suggest [42]. This highlights the potential role of prospective studies in exploring outcomes and euthanasia rates and reasons in haemangiosarcoma cases. Whilst it was not within the scope of this project to explore factors motivating owners to pursue euthanasia, it is likely that it was not only related to the tumour aggressive behaviour, but also other factors such as the severity of the clinical signs and perceived acute deterioration in quality of life, together with the age at presentation (adult to elderly for large breeds), probable presence of age-related comorbidities and treatment costs.

Future work could also attempt to identify a control population of other splenic tumours to evaluate whether the presumed haemangiosarcoma group truly has distinct survival and could improve the external validity of the study. However doing this has some challenges that would need to be taken into consideration: other neoplasms within the spleen are rare (only account for 25% of malignant splenic tumours) and include (aside from lymphoma) mainly other sarcomas (histiocytic sarcoma, leiomyosarcoma, fibrosarcoma, undifferentiated sarcomas and liposarcoma) for which survival data in the literature is scarce [4]. Additionally, histologically these sarcomas can often be challenging to distinguish from haemangiosarcoma and immunohistochemistry may be needed to achieve a definitive diagnosis, which may not be commonly pursued in the general practice setting.