We developed a model to predict the PoD of patients with advanced cancer receiving palliative homecare in Iran, as a representative of the countries with “isolated palliative care provision” (aka group 3a countries based on WHO categorization of the palliative care development [4]). We identified age, tumor topography, number of homecare visits in the last two weeks, number of CTPs in the last two weeks, and the site of registration as predictors of the PoD. In tumor-stratified subgroup analysis, models specific to hematopoietic and gynecologic cancers showed the highest accuracy and in site-specific subgroup analyses, the Tehran model had higher accuracy than Isfahan. The subgroup analyses for sex and marital status did not enhance the accuracy of prediction (Fig. 3C and G). We identified determinants of PoD among all three categories of variables suggested by Gomes and Higginson [11].

Similar to previous studies [34,35,36], our study identifies hematopoietic malignancies as a predictor of death at the hospital. Though previous studies attributed this to the fewer admissions of these patients to palliative care services [11], we identified hematopoietic cancers as a predictor of death at the hospital among the patients who were enrolled in a homecare palliative care program. Thus, we suggest that the higher likelihood of death at the hospital among these patients might be due to the different trajectories of hematopoietic malignancies [37]. The vicissitudinous nature of the hematological cancers is potentially associated with abrupt changes in patient’s situation, emergency admissions to hospital and prolonged hospitalization and sometimes, unpredictable death amid curative treatment period which collectively highlight the need for tailored end-of-life care for these patients. Likewise, our results show that gynecological cancers have a higher likelihood of death at the hospital. Although the incidence of gynecological cancers is limited to women, the observed difference may not be surrogated by sex differences since in our analyses, sex is not significantly associated with the PoD. A recent study by Kobo et al. that relies on data from circa 6 million American cancer patients points out that malignancy-related issues are the most common cause of hospital admissions among cancer patients [38], thus, future studies should consider the incorporation of cancer-related complications, that may differ by the types of malignancies, in prediction models of PoD.

While in our study the number of comorbidities was not associated with the PoD, Izquierdo-Porrera et al. identified the digestive comorbidities at the time of admission as a predictor of death at the hospital among men [39]. In addition, Kobo et al. identified infectious causes followed by gastrointestinal and cardiovascular causes as the most common non-malignancy-related causes of hospital admission among cancer patients [38]. Detailed coverage of chronic and acute health complications in medical records, utilizing standard recording systems like ICD-11 [40], and incorporating such data in future predictive models presumably fills this gap.

Among the individual factors we analyzed, the patient’s gender and income quartile were not associated with the PoD and the higher age of the patient was a predictor of death at home. Gomes and Higginson highlighted 16 studies that inconsistently reported the predictive ability of age for PoD in their 2006 systematic review [11]. The predictive value of age has remained inconsistent in recent studies [13, 15]. We speculate that the significance of age as a predictor of PoD is linked to the local sociodemographic factors as well as available resources for elderly care, and thus may vary in different communities.

A recent study by Fereidouni et al. reported that over 75% of a group of Iranian advanced cancer patients preferred to die at home [25]. The concordance between the patient’s preferred PoD and the actual PoD is an important indicator of the effectiveness of palliative care services [41], though, we did not include patients’ preferences in our study. Identification of the patient’s preferred PoD is hindered by the lack of legal and executive infrastructure of advanced directives in Iran and many other countries [42]. While this calls for the necessity of global implementation of advanced directives, on the other hand, a study from Korea indicated that nearly 70% of cancer patients who preferred to die at home have died in the hospital [14] suggesting that the identification of the preferred PoD cannot predict PoD per se.

In all models we generated, the prediction ability of the number of physician home visits in the last two weeks was statistically significant with the highest impact on the PoD. This can result from the potential ability of homecare visits to respond to the patient’s needs at home in the final stage of life. This finding is in accord with the findings from earlier studies indicating that dying at home is correlated with ≥ 3 nurse home visits during the first week of homecare, ≥ 5 nursing home visits in the final week of life [16], the overall number of homecare visits [19] and overall number of physician visits [13]. Despite this, the total number of home visits, the number of visits by other medical professionals [13], the frequency of homecare visits in the final month of life [15] and the number of nurse visits [10, 13] have not contributed to the prediction of the PoD. The inconsistency between the specifications of services and the measures for intensity of care may have contributed to the different conclusions obtained in different studies. We also analyzed the effect of CTP and CFP on the PoD determination. While the number of CFPs was not associated with the PoD, unexpectedly the higher number of CTPs in the final two weeks of life was shown to be associated with an increased likelihood of death at the hospital.

We identified different powers of prediction of the main model in different sites of services. The site-specific subgroup models also differently weighted the predictors of death. This difference may reflect the variations in local practice, local standard of care, socio-geographical background, host community, and local access to other health care services. A study from Italy identified a significant variation in the proportion of deaths that occurred at home among 13 provinces where the study was performed, ranging from 31.4 to 73.3%. Differences in perception of dying at home among different cultural groups and inappropriate utilization of hospital services in different areas are suggested as the factors that underlie this finding [43]. Another study performed in Nova Scotia, Canada, indicates that patients residing in the metropolitan Halifax region were more likely to die at home than those living in other regions of the province. The article speculates that the difference in care culture and the public access to community-based services in different regions have led to this contrast [44]. Comparably, Temkin-Greener and Mukamel revealed that the prediction of the PoD varies across 12 different sites of the Program of All-Inclusive Care for the Elderly (PACE). The authors suggested the variation among PACE services and the local facilities as the reasons for the observed variation [18]. In this study, the distribution of demographic variables was not significantly different between the 2 sites of MACSA services (Supplementary Table 1). Thus, the difference in the proportion of home deaths in different sites rather seems to be - at least in part - a result of differences in sociological specifications and service-related variables (Supplementary Table 1). However, this is not limited to the variables explored in this study. These findings imply the importance of developing local models, incorporating social, cultural, logistic, and demographic specifications linked to each geographical location, especially for the countries in the preliminary stages of palliative care development.

Our study identified no correlation between income quartile, residency status, number of co-residing family members, presence of CRFM < 18 years, and the average age difference between patients and CRFM. Consonantly a study by Masucci et al. found that the caregiver’s age or gender is not associated with the PoD [10]. However, the caregiver’s age ≥ 55 years was shown to be connected with the patient’s death at home in the model developed by Tay et al. [12]. The presence of caregivers [18] or co-residents [10] is shown to be a predictor of death at home in previous studies, a variable that is not measured in our study. Furthermore, marital status, which previously was shown to be a predictive factor of the PoD, did not emerge as a significant predictor of the PoD in this study.

To the best of our knowledge, this study is the first effort to develop a prediction model for PoD in countries with isolated provisions of palliative care, where the majority of annual human deaths occur [4]. This study is also the first of its kind in developing countries and also the west of Asia. This study benefits from a large sample size and the data collected during the early years of the development of a donation-based homecare cancer palliative service, which can give insight into the predictors of death in the early years of implementing palliative care in other similar countries. Another advantage of this work over past studies was the incorporation of the telemedicine-related variables, CTP and CFP to the prediction model.

Another point needs to be discussed is the possible effects of COVID-19 pandemic on changing patterns of death at home or hospital in cancer patients. Based on Turtle, L., et al., during the pandemic the cancer patients with COVID-19 have had a higher rate of death at hospital compared to the pre-pandemic era. Thus, it is tempting to speculate that the COVID-19 diagnosis could has served as a predictor for hospital death during the pandemic [45]. We could not find any article reporting changes in the health care system or cancer patients’ needs in the post-pandemic era except that the incidence of some cancers has increased due to COVID-19 infection, however, it is unlikely that the care outcome and place of death has been affected by the changes in incidence since the overall change the incidence is not very large. To shed light on this question the same team is researching the changes in the incidence of symptoms experienced by cancer patients before, during COVID-19 and after COVID-19.

Our study was limited by the number of measured variables, especially social and psychological variables (including patient’s and caregivers’ preferred PoD, the cultural determinants of PoD, intimacy and emotional connection with the family and having independence in doing daily tasks at the end of life are among the psychological factors potentially affecting the preferred place of death), geographical variables (for example, patients’ home distance from the health care centers) and patient’s disease profile (type of comorbidities, symptoms and end of life complications). Overall, we considered many potential predictors of PoD, however as mentioned, we did not have access to some other important variables which led to diminishing the efficacy of our fitted prediction model reflected in its not so high sensitivity and specificity. Missing data in co-residency variables and income groups were other limitations. Our model does not distinguish hospital referrals at short intervals before death (e.g., < 24 h.) from longer periods of hospital admissions that concluded with death nor hospital referral by palliative care physicians. While this may challenge the number of visits in last two weeks of life as a predictive variable, this should be considered that our model identifies the effect of the intensity of service-related variables in last two weeks of life on PoD, regardless of the underlying mechanisms.

Some limitations arise from the fundamental complexity of research about death. While factors like cultural perspective on death and dying, the tendency to talk about death and communicating wishes, subcultures, and legal protocols can affect the patient’s PoD, they are usually overlooked in studies that aim to develop a prediction model. The possible reason for such ignorance could be the complexity of measuring such variables. The other hurdle might be the difficulty of handling multiple questionnaires by patients, especially those in the final stages of their lives.