Longitudinal symptom network patterns in cancer patients during the end of life

Abstract

Patients with cancer at the end of life experience a wide range of distressing physical and psychological symptoms, which often interact and intensify each other. Understanding how these symptoms are connected is essential for improving quality of life and providing effective palliative care. This study analyzed data from 26,318 medical visits of 8,026 patients receiving palliative care, focusing on symptom interactions during the final months of life. Symptom networks were estimated using the Ising Model across six time periods prior to death, with centrality measures used to identify key symptoms, and the Walktrap algorithm applied to detect symptoms communities. The results revealed that different symptoms and their relationships were important at different stages, although certain symptoms—such as weakness, cognitive problems, and issues with physical appearance—consistently played central roles. Strong and persistent connections were observed between nausea and vomiting, as well as anxiety and depression. Additionally, five distinct symptom communities were identified over time. These findings provide insight into the dynamic nature of symptom interactions near the end of life and highlight key targets for timely and tailored palliative interventions, ultimately aiming to improve patient comfort and support a more peaceful dying process.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-29627-6.

Introduction

Cancer is one of the leading causes of death in the world¹. Approximately 20 million new cases of various types of cancer were diagnosed and, nearly 10 million deaths happened because of the disease². Cancer, with its diverse forms and complex symptoms, remains one of the most challenging issues in modern healthcare³. Several factors influence the quality of life of cancer patients⁴. Among them, cancer-related symptoms play a crucial role in decreasing the quality of life and directly impact patients’ well-being⁵. While cancer patients often experience multiple symptoms simultaneously, most previous studies have focused on one or a few symptoms. However, symptoms rarely occur in isolation. They often co-occur and interact, forming dynamic patterns that may influence patient outcomes more strongly than single symptoms alone^6–8. This limited focus leads to a lack of comprehensive understanding of the overall symptom burden^8,9.

Network analysis (NA) provides a promising approach to capture these interactions, highlighting both clusters of symptoms and central symptoms that may play a key role in overall burden. NA is a methodology that focuses on understanding the nature of relationships between elements (nodes) in a system and how these relationships (edges) influence each other^10,11.

Previous studies have applied network approaches to cancer populations, but most have focused on cross-sectional data or specific cancer types. Evidence on symptom networks during the last year of life remains sparse, despite its critical importance for clinical care and decision-making^8,12,13. For patients at the end of life (EOL), symptoms often become more severe, more numerous, and more interconnected, creating complex patterns that can overwhelm both patients and caregivers^14–16. Understanding these symptom dynamics specifically in EOL populations is essential, as it can guide timely interventions, improve comfort, and support personalized palliative strategies^15,17,18. Moreover, existing work is often based on small samples, limiting generalizability.

To the best of our knowledge, none of the existing studies have focused on cancer symptoms in EOL patients with the mentioned approach. Therefore, the present study aims to address this research gap by answering the following questions: (1) How does the structure of symptom networks in EOL patients inform our understanding of symptom interactions? and (2) which symptoms are more influential and play a central role in the network?

Literature review

To provide an overview of previous research, we conducted a narrative literature review focusing on studies relevant to symptom prevalence and network analysis in cancer patients. Articles were selected based on their relevance to cancer symptom patterns, symptom networks, and longitudinal or cross-sectional analyses. We prioritized peer-reviewed studies that examined multiple cancer types, specific populations, or EOL patients to highlight gaps in existing knowledge.

Previous studies have shown that cancer patients suffer from various psychological and physical symptoms depending on the type of treatment and disease^19,20. If these distressing symptoms are not managed effectively, they can significantly affect patients’ quality of life^5,21,22. Comparative studies have found that cancer patients not only experience more symptoms than non-cancer individuals but also report lower well-being scores²³. Therefore, it is essential to focus not only on increasing the life expectancy of cancer patients but also on enhancing their quality of life during this extended lifespan²¹.

Previous studies assess symptoms individually and often neglect the impact of symptoms on each other^6,8,24,25. However, due to the complex nature of cancer symptoms, it is necessary to explore them using new approaches²⁶. One novel method is network analysis, which is a powerful data-driven approach that can be used to uncover the complex and hidden relationships between symptoms^11,13,27,28. This approach helps us identify meaningful connections between symptoms and provides better disease interpretation by identifying key nodes (core symptoms)^9,29. There are two ways to conduct research on cancer symptoms based on network analysis: (1) researchers select their study sample size based on a specific type of cancer, and (2) the study sample includes a set of different cancer types⁹.

One of the studies focusing on a specific type of cancer is the research by Röttgering et al., who examined symptom networks in glioma patients across different network subgroups (surgery, fatigue, and tumor grade)¹³. Their findings revealed that, in all networks, severe fatigue, depression, and social functioning were the most interconnected symptoms. Additionally, symptom connections were stronger in the fatigue network compared to non-fatigue patients, highlighting the role of fatigue in the overall symptom experience of glioma patients.

In another study, Jing et al.³⁰ investigated the symptom network of breast cancer patients undergoing hormone therapy. They found that irritability, mood swings, joint pain, and decreased sexual interest were key symptoms in this group of patients. Similarly, Zeng et al.²⁹ examined symptom networks in multiple myeloma patients undergoing chemotherapy. Their network analysis showed that pain was the most common symptom, while worry had the greatest impact on other symptoms. A study on liver cancer patients²⁸ also revealed that distress was the central symptom with the highest strength and betweenness centrality, while fatigue showed the highest closeness centrality. As nausea was identified as the strongest bridge symptom, the authors suggested that therapeutic interventions for this symptom could lead to better symptom management for this group of patients.

While the above studies examined symptom networks by focusing on a single type of cancer, other studies have investigated multiple cancer types to explore symptom clusters. For example, De Rooij et al.⁸ aimed to identify symptom clusters and assess how these clusters vary across different cancer types, treatment modalities, and duration (short-term vs. long-term). They found that fatigue was the most central symptom, linked to emotional symptoms, cognitive symptoms, pain, shortness of breath, and appetite loss. Similarly, Zhu et al.⁶ analyzed symptom networks across various cancer types based on cancer duration. Their findings revealed that network density significantly differed according to the length of cancer experience. Symptoms such as distress, sadness, and appetite loss were more prominent in patients with a shorter duration of cancer.

Some studies, like those mentioned above, focused on both physical and psychological symptoms, while others specifically targeted psychological symptoms and explored their underlying mechanisms. Recently, psychological symptoms such as depression have been conceptualized as a system³¹. In this approach, symptoms are viewed as interacting and influencing each other rather than being isolated indicators of an underlying structure³². For example, Sharpley et al.³³ investigated depressive symptoms in prostate cancer patients. Their network analysis revealed that anhedonia (loss of pleasure) was the most central symptom. Sleep disturbances, fatigue, suicidal thoughts, feelings of worthlessness, and depressed mood were among the symptoms closely related to each other and to anhedonia. In comparison to the general population, Hartung et al.²⁴ found that cancer patients experienced more severe depressive symptoms. However, the network of symptom relationships in cancer patients was less interconnected than that of the general population. This indicates that cancer patients have different needs compared to the general population. Bickel et al.³² in a similar study, examined the interactions between depressive symptoms in both Contemporaneous and temporal networks while also considering the care needs of cancer patients. Their findings revealed that cancer patients want more support for fatigue, feelings of weakness, lack of pleasure, and sleep disturbances. Bickel et al.‘s study highlighted some of the needs of cancer patients and the relationships between their symptoms. However, in symptom network analysis, it is crucial to explore various populations, as different groups may have unique needs³⁴.

Therefore, examining the needs and symptoms of patients across different stages or age groups can lead to more personalized treatments and better responses to the specific needs of cancer patients. In this context, the study by Fang et al.³⁵ on children with acute leukemia undergoing chemotherapy revealed that lack of energy, pain, hair loss, sadness, and worry were common symptoms in this population. Among these, worry and irritability were the central symptoms, indicating that emotional symptoms significantly influenced the symptom network of children. Similarly, the study by Kuang et al.³⁶ on elderly cancer patients found that fatigue, sleep disturbances, and memory problems were the most common and severe symptoms in this group. Overall, the symptom network density was higher in elderly patients with shorter survival times and comorbidities.

Most of the mentioned studies examined symptoms cross-sectionally. However, longitudinal analysis of symptom networks can provide deeper insights. For instance, Kalantari et al.⁹ investigated the symptoms of patients with gastrointestinal, breast, gynecological, and lung cancers at six-time points throughout two chemotherapy cycles. Additionally, they separately analyzed symptoms related to breast cancer. Their findings revealed that symptoms and their interconnections changed throughout the chemotherapy cycle and varied depending on the specific type of cancer. Similarly, Shim et al.³⁷ examined symptom networks at three-time points (before surgery T0, one week after surgery T1, and 3 to 6 months after surgery T2) in patients with gastric cancer. Their results showed that distress and sadness were central nodes across all networks. Anxiety was closely related to emotional and physical health at T1, while depression was linked to functional health at T0 and T2. A summary of the studies is presented in Table 1.

Table 1.

A summary of previous studies.

Authors (year)	Focus of Study	Longitudinal data	No of participants		No of symptoms	Treatment type	Cancer type
(Papachristou et al., 2019)	Multidimensional Symptom Network Analysis in Oncology	N/A	1328		32 + 6	Chemotherapy	Breast, gastrointestinal, gynecological, lung
(Zhu et al., 2023)	Contemporaneous Symptom Network Analysis of Multidimensional Symptom Experiences in Cancer Survivors	N/A	1065		13	Combination of treatments	Different types of cancer
(de Rooij et al., 2021)	Network Analysis for Symptom Clustering in 1,330 Patients Across 7 Cancer Types	N/A	1330		30	Surgery, radiotherapy, chemotherapy	Colorectal, Breast, Ovarian, Thyroid, Chronic Lymphocytic Leukemia, Hodgkin Lymphoma, and Non-Hodgkin Lymphoma
(Jing et al., 2023)	Symptom Network Analysis During Endocrine Treatment in Breast Cancer Patients	N/A	613		19	Hormone therapy	Breast
(Hartung et al., 2019)	Network Analysis of Depressive Symptoms in Cancer Patients Compared to the General Population	N/A	4020 cancer patients compared to 4020 non-cancer		9	N/A	Different types of cancer
(Sharpley et al., 2023)	Network Analysis of Depression in Prostate Cancer Patients	N/A	555		9	Hormone therapy, surgery, radiotherapy, other	Prostate
(Röttgering et al., 2023)	Multidimensional Understanding of Symptoms and Quality of Life through Symptom Network Analysis in Glioma Patients	N/A	256		N/A	Surgery	Glioma
(Bickel et al., 2022)	Dynamic Symptom Network Structure Analysis of Depression in Cancer Survivors and Their Preferences for Psychological Care	Yes	52		8	Combination of treatments	N/A
(Zeng et al., 2023)	Network Analysis of Symptoms in Multiple Myeloma Patients Undergoing Chemotherapy	N/A	177		32	Chemotherapy	Myeloma
(Xu et al., 2024)	Symptoms experienced after transcatheter arterial chemoembolization in patients with liver cancer	N/A	1207		13	Chemoembolization	Liver
(Kuang et al., 2024)	Symptom network analysis in older adults with cancer	No	485		13	Surgery, chemotherapy, radiotherapy	Gastrointestinal, Breast, Lung, Urinary
(Fang et al., 2023)	Symptom Network Analysis in Children Aged 5 to 18 with Cancer	No	469		31	Chemotherapy	Acute Leukemia
(Kalantari et al., 2022)	Network Analysis of Symptoms at Six Time Points During Two Phases of Chemotherapy	Yes	987		32 + 6	Chemotherapy	Breast, gastrointestinal, gynecological, lung
(Shim et al., 2021)	Symptom Analysis of Gastric Cancer Patients Before and After Surgery	Yes	256		13	Surgery	Gastric

Considering the explanations provided about previous studies and the various populations and characteristics examined, it is clear that certain gaps exist that have not been explored or have received limited attention in the context of cancer symptoms. One of the most significant gaps is the lack of research on symptoms in cancer patients EOL. Most previous studies excluded patients who were in the final stages without tumor treatment options. This gap is noteworthy because cancer patients in the end stages require special care and effective symptom management. According to the authors’ knowledge, this study is the first to investigate symptom relationships in this patient group using a network analysis approach. Also, this study aims to explore symptom networks longitudinally across six periods. While most of the previous studies conducted cross-sectional research, longitudinal research improves understanding and extracts deeper insights from networks. Additionally, this study utilizes a large sample of patients and their visits to estimate symptom networks. In contrast, most previous studies have relied on relatively small sample sizes. This large sample size allows for greater generalizability of findings and enhances the accuracy of network estimation.

Methodology

Data collection

We analyzed electronic health record (EHR) data from the MACSA project, which integrates palliative care information from multiple centers from March 2017 to March 2024. Symptom-related data are gathered through questionnaires. These questionnaires are completed by trained healthcare professionals, including nurses and physicians, during home care visits. The dataset included 26,318 visits from 8,026 patients who subsequently died of cancer. For analysis, we focused exclusively on the final years of life. To examine changes over time, visits were grouped into six periods preceding death.

Study population

The target population for this study consists of cancer patients who received home care services and were in the final stage of life. The inclusion criteria are as follows: (1) diagnosed with cancer; (2) 18 years of age or older; (3) having a death date record; and (4) having at least one documented medical visit before death. Note that this study does not include any living participants.

Symptom measures

The data used in this study include demographic variables, clinical variables and symptom-related data. The demographic data include marital status, age, and gender. Clinical data include tumor type, tumor grade, and disease duration. At each visit, clinicians assessed 37 binary symptom items (0 = absent, 1 = present). These included common physical and psychological symptoms such as pain, fatigue, dyspnea, anxiety, and depression. The full list of symptoms and their coding is provided in Supplementary Table S1.

Visits management

Based on previous studies^14,38,39 that conducting a longitudinal study, the visits in the present research were categorized into six-time periods before death: (1) one month before death (T1), (2) one to two months before death (T2), (3) two to three months before death (T3), (4) three to six months before death (T4), (5) six to twelve months before death (T5), and (6) more than twelve months before death (T6). To calculate these time periods, the time interval between each visit and the death date was calculated in days⁴⁰. The first period covered days 0 to 30 (day 0 indicates that the patient passed away on the visit day). Similarly, the second period covered days 31 to 60 and so on.

If a patient had only one visit within a given time period, that visit was assigned to the corresponding interval. If a patient had multiple visits within the same period, the data from those visits were aggregated and treated as a single visit for that interval.

Network estimation

All networks were visualized using RStudio version 4.4.1. The common method for estimating symptom networks is the Pairwise Markov Random Fields (PMRF)^9,41. When the data are binary, the Ising model is a suitable alternative to PMRF. To build the Ising model, the IsingFit package in R was used. This method combines L1-regularized logistic regression with model selection based on the Extended Bayesian Information Criterion (EBIC). The hyperparameter gamma was set to 0.5 for all time periods to create sparse networks and avoid spurious edges. The qgraph package was also used to visualize the networks. In our symptom networks, each symptom was considered a node, and edges represented conditional independent relationships between them⁴². Additionally, the networks are weighted, with thicker edges indicating stronger connections. Green edges represented positive connections while red edges represented negative connections.

Node centrality and community detection

To identify key symptoms, we computed centrality indices for each network. In cancer symptom networks, the most common centrality measures are betweenness, closeness, and strength centrality^27,32,33,37. Betweenness centrality represents the number of shortest paths that pass through a node. Nodes with high betweenness centrality act as bridges in the network. Closeness centrality indicates the average distance of a symptom from all other symptoms in the network. A higher closeness centrality means that the symptom is, on average, closer to all other symptoms. Strength centrality reflects the number of direct connections a node has with other nodes. This measure shows the direct influence of symptoms on each other^13,30. Also Community detection was examined using the Walktrap algorithm, which detects groups of symptoms that frequently co-occur⁹.

Accuracy and stability estimation

The accuracy and stability of the networks at each time period were assessed using the Bootnet package in R. To evaluate edges’ accuracy, bootstrapping was performed with nBoots = 1000, using 95% bootstrap confidence intervals (CI). Furthermore, a case-dropping bootstrap was conducted to estimate the centrality stability (CS) coefficient. If the CS coefficient is greater than 0.25, it indicates that the centrality metrics are reliable, and a CS coefficient of more than 0.50 is preferable⁴³.

Ethics approval and consent to participate

The study was conducted in accordance with relevant guidelines and regulations. The original data were collected by the Iranian Cancer Control Center (MACSA) under ethical approval from its institutional review committee, and informed consent was obtained from all participants prior to death. The present research involved secondary analysis of this de-identified dataset and was reviewed and approved by MACSA for research use. No direct patient contact occurred, and all data were anonymized prior to analysis.

Result

Sociodemographic and clinical characteristics of participants

In this study, 26,318 visits recorded for 8,026 cancer patients were analyzed. The demographic results showed that 53% of the patients were male, and the average age of the patients was 63 years. Regarding relationship status, approximately 75% of the patients were married, 16% had lost a spouse, 5% were single, 3% were divorced or separated, and nearly 1% had no recorded status. Clinical results revealed that, on average, patients had been battling the disease for nearly 2.5 years. The duration of cancer involvement for 6,230 patients (approximately 85%) was less than 5 years, while for 1,130 patients (approximately 15%) it was over 5 years, and for 666 patients, the disease duration was unknown. In terms of cancer tissue type, gastrointestinal cancers were the most common, accounting for 38.6% (2,974 cases). Following that, bone and soft tissue cancers accounted for 14.2% (1,095 cases), and breast cancer accounted for 12% (929 cases). Cancers of the female reproductive system made up 5.9% (458 cases), brain cancer 5.7% (445 cases), urinary system cancers 4.8% (371 cases), and male reproductive system cancers 3.7% (289 cases). Additionally, 4.6% (359 cases) of cancers had an unknown origin, and 9.5% (776 cases) were classified as other cancers. The type of cancer was not recorded for 330 patients. In addition to the tumor types, the data related to tumor grading were also examined in this study. The results showed that 15.5% (1,238 cases) of tumors were grade three, 13% (1,049 cases) were unknown or grade X, 10% (808 cases) were grade two, 7% (553 cases) were grade one, and 4.2% (337 cases) were grade four. A high percentage of patients (50.3%, equivalent to 4,041 cases) lacked registered information on tumor grading.

Prevalence of symptoms

The analysis was conducted across six defined time periods preceding death. As shown in Table 2, the most prevalent symptoms were pain (59% mean prevalence), appetite problems (55%), and fatigue (42%). These physical symptoms were consistently common throughout the final year of life. Alongside them, several psychological symptoms were also prominent, including irritability (37%), depression (32%), and anxiety (29%). The relatively high frequency of these psychological manifestations highlights their clinical importance and suggests that they should be considered as significant as physical symptoms in the context of EOL care.

Table 2.

Top prevalent symptoms.

Symptoms	Mean prevalence	First period	Second period	Third period	Fourth period	Fifth period	Sixth period
Pain	%59	%57	%54	%51	%58	%60	%74
Appetite problems	%55/33	%59	%56	%53	%55	%53	%56
Fatigue	%42/50	%42	%36	%31	%38	%42	%66
Constipation	%38	%41	%38	%34	%39	%36	%40
Irritability	%37/67	%30	%31	%32	%40	%45	%48
Sleep problems	%33	%29	%30	%27	%34	%37	%41
Depression	%32/50	%23	%26	%30	%36	%39	%41
consciousness impairment	%29/83	%45	%35	%28	%30	%23	%18
Anxiety	%29/67	%21	%22	%24	%31	%35	%45
Breathing problems	%29/17	%36	%28	%23	%26	%25	%37

Although prevalence rates showed some fluctuations across time, notable trends emerged. Pain and fatigue increased as death approached (especially 3 months prior to death), whereas appetite problems remained relatively stable. By contrast, consciousness impairment displayed a distinct temporal pattern, rising from 18% in patients more than one year before death (sixth period) to 45% in the last month of life (first period). This progressive increase suggests a worsening trajectory of neurological decline near the end of life. Collectively, these findings demonstrate that both physical and psychological symptoms play a central role in the burden experienced by patients and must be systematically addressed in palliative care planning (compelete prevalence of all symptoms is reported in Supplementary Table S2).

Prevalence of visits

The total number of visits recorded before aggregation was 26,318. Among these, there were 8,957 visits in the first period (34.03%), 4,246 visits in the second period (16.13%), 2,460 visits in the third period (9.35%), 3,972 visits in the fourth period (15.09%), 3,308 visits in the fifth period (12.57%), and 3,375 visits in the sixth period (12.82%). These visits were attributed to a total of 5,510 patients in the first period, 2,819 patients in the second period, 1,734 patients in the third period, 1,945 patients in the fourth period, 1,303 patients in the fifth period, and 963 patients in the sixth period. Visits were recorded for patients in the following ways: 3,169 patients were recorded in only the first period; 1,678 patients were recorded in both the first and second periods; 671 patients were recorded from the first to the third period; 416 patients were recorded from the first to the fourth period; 219 patients were recorded from the first to the fifth period; and 119 patients were recorded in all six periods. During the sixth period, a total of 3,375 visits were recorded. Of these, 2,357 visits (69.8%) occurred between 365 and 730 days (one to two years) before death. Additionally, 609 visits (18%) were noted two to three years before death, while 409 visits (12%) were recorded three to six years prior to death.

Symptom network analysis

This section presents the results of the symptom network analysis across six time periods before death. For each period, we estimated the network structure, examined centrality indices, and identified symptom communities using the Walktrap algorithm. To improve readability, detailed network plots and centrality diagrams have been moved to the Supplementary Figures (S1–S12).

Core symptoms and their relations in period 6 (i.e., more than 12 months before death)

In this period, the estimated network included 31 symptoms connected by 66 edges, with six symptoms remaining isolated (diarrhea, seizures, weakness, obsession, tinnitus, and taste changes). Most associations were positive, with the exception of a negative edge between physical appearance issues and consciousness impairment. The strongest connections were observed between clinically related symptoms such as blurred vision– bleeding in various parts of the body, vomiting – nausea, feeling of anger – dehydration, irritability – anxiety and depression – anxiety.

Centrality analysis indicated that urinary problems emerged as the most central symptom, showing the highest strength, closeness, and betweenness values. Dehydration, constipation and sleep problems were also highly connected. Community detection in Period 6 revealed six clusters, ranging from small and symptom-specific (e.g., nausea–vomiting) to large and systemic (e.g., sleep, breathing, and appetite problems). A summary of the communities is presented in Table 3.

Table 3.

Symptom communities of the sixth period.

Community	Main theme (Interpretation)	Symptoms included
1	Psychological	Irritability, Depression, Anxiety
2	Gastrointestinal	Vomiting, Nausea
3	Neurological	Dizziness, Balance disorders
4	Mixed somatic	Weight loss, Blurred vision, Anger, Dehydration, Bleeding
5	Pain & cognitive-related	Pain, Cognitive problems, Fatigue, Gastrointestinal problems, Itching, Appearance issues
6	Broad systemic	Sleep problems, Breathing problems, Urinary problems, Swallowing difficulties, Consciousness impairment, Appetite issues, Cough, Bedsores, Sweating, Constipation, Oral ulcers/inflammation, Pallor, Edema

Core symptoms and their relations in period 5 (i.e., 6–12 months before death)

In this period, the network consisted of 30 symptoms connected by 77 edges, with seven isolated symptoms (sweating, seizures, obsession, tinnitus, anger, taste changes, and bleeding). Most associations were positive, with the exception of two negative edges (consciousness impairment–anxiety and pain–weight loss). Strongest connections were observed between clinically related symptoms, such as vomiting–nausea, irritability–anxiety, and depression–anxiety, as well as consciousness impairment with both weakness and swallowing difficulties. Centrality analysis revealed that consciousness impairment was the most central symptom across all indices. Nausea and anxiety were also highly connected (strength centrality), while fatigue and appearance issues had higher betweenness values, suggesting a bridging role. Fatigue and weakness additionally showed high closeness centrality. Community detection identified six symptom clusters, summarized in Table 4.

Table 4.

Symptom communities of the fifth period.

Community	Main theme (Interpretation)	Symptoms included
1	Gastrointestinal & appetite	Pain, Appetite problems, Weight loss, Vomiting, Nausea, Oral ulcers/inflammation, Constipation
2	Psychological	Irritability, Depression, Anxiety
3	Respiratory	Breathing problems, Cough
4	Neurological (sensory)	Dizziness, Blurred vision
5	Fatigue & somatic issues	Fatigue, Gastrointestinal problems, Itching, Diarrhea, Appearance issues
6	Broad systemic	Sleep problems, Cognitive problems, Swallowing disorders, Urinary problems, Consciousness impairment, Bedsores, Balance disorders, Weakness, Pallor, Edema

Core symptoms and their relations in period 4 (i.e., 3–6 months before death)

In this period, the network consisted of 31 symptoms connected by 99 edges, with six symptoms (sweating, diarrhea, obsession, blurred vision, anger, and bleeding) appearing as isolated nodes. Unlike earlier periods, several negative associations were detected, including pain–weakness, pain–weight loss, and consciousness impairment–anxiety. The strongest positive connections were observed between clinically related symptoms such as weakness- tinnitus, vomiting - nausea, weakness - changes in taste perception, fatigue - weakness, balance disorders - sweating, and depression - anxiety. Centrality analysis highlighted weakness as the most central symptom across all indices (strength, betweenness, and closeness). Physical appearance issues and consciousness impairment also ranked highly in strength centrality, while pain (betweenness) and balance disorders (closeness) emerged as additional influential nodes. Community detection identified six distinct symptom clusters, summarized in Table 5.

Table 5.

Symptom communities of the forth period.

Community	Main theme (Interpretation)	Symptoms included
1	Systemic & functional	Sleep problems, Respiratory issues, Cognitive problems, Urinary problems, Swallowing disorders, Consciousness impairment, Cough, Bedsores, Constipation, Seizures, Oral ulcers/inflammation, Pallor, Dehydration, Edema
2	Physical/appearance-related	Pain, Weight loss, Fatigue, Gastrointestinal problems, Appetite issues, Itching, Appearance issues
3	Neurological (balance)	Dizziness, Balance disorders
4	Gastrointestinal	Vomiting, Nausea
5	Neurological (weakness & sensory)	Weakness, Taste changes, Tinnitus
6	Psychological	Irritability, Depression, Anxiety

Core symptoms and their relations in period 3 (i.e., 2–3 months before death)

In this period, the network included 30 symptoms connected by 75 edges, with seven symptoms (sweating, dizziness, tinnitus, anger, bleeding, blurred vision, and obsession) isolated. A relatively high number of negative associations (n = 12) were detected, most notably between pain and weakness, pain and weight loss, and pain and anxiety. Positive edges remained strongest between clinically related symptoms such as weakness - changes in taste perception, vomiting - nausea, fatigue - weakness, balance disorders - weakness, and irritability – weakness. Centrality analysis identified weakness as the most influential symptom, with high values across strength, betweenness, and closeness. Physical appearance issues and gastrointestinal problems also ranked highly in strength, while pallor, fatigue, and appearance issues played important bridging roles. Community detection revealed seven clusters, summarized in Table 6.

Table 6.

Symptom communities of the third period.

Community	Main theme (Interpretation)	Symptoms included
1	Systemic signs	Pallor, Dehydration, Edema
2	Neurological/functional	Cognitive problems, Swallowing disorders, Consciousness impairment, Bedsores
3	Pain & somatic burden	Pain, Fatigue, Gastrointestinal problems, Itching, Diarrhea, Appearance issues, Oral ulcers/inflammation
4	Respiratory/sleep	Sleep problems, Respiratory issues, Cough
5	Nutritional	Weight loss, Appetite issues, Constipation
6	Gastrointestinal	Vomiting, Nausea
7	Neurological (motor/sensory)	Balance disorders, Weakness, Taste changes

Core symptoms and their relations in period 2 (i.e., 1–2 months before death)

In this period, the network comprised 29 symptoms connected by 114 edges, with eight isolated symptoms (bedsores, sweating, diarrhea, taste changes, bleeding, tinnitus, obsession, and anger). Fourteen negative associations were detected, particularly involving pain and weakness, pain and weight loss, and consciousness impairment with gastrointestinal problems. Positive associations remained strong between clinically related symptoms, such as vomiting - nausea, fatigue - weakness, depression - anxiety, fatigue - physical appearance issues and balance disorders – weakness. Centrality analysis identified physical appearance issues, fatigue, and weakness as the most central symptoms across strength, betweenness, and closeness indices, indicating their prominent roles in this stage of illness. Community detection revealed six clusters, summarized in Table 7.

Table 7.

Symptom communities of the second period.

Community	Main theme (Interpretation)	Symptoms included
1	Gastrointestinal	Vomiting, Nausea
2	Neurological (sensory)	Seizures, Blurred vision
3	Neurological (balance)	Dizziness, Balance disorders
4	Psychological	Irritability, Anxiety, Depression
5	Pain & somatic burden	Pain, Fatigue, Gastrointestinal problems, Itching, Weakness, Appearance issues
6	Broad systemic	Sleep problems, Respiratory issues, Weight loss, Consciousness impairment, Urinary problems, Swallowing difficulties, Appetite issues, Cough, Cognitive problems, Constipation, Pallor, Oral ulcers/inflammation, Dehydration, Edema

Core symptoms and their relations in period 1 (i.e., 1 month before death)

In the final month of life, the network included 33 symptoms and 179 edges, with four symptoms (sweating, tinnitus, obsession, and bleeding) isolated. This period exhibited the highest number of negative associations (n = 24), particularly involving pain, weakness, appearance issues, and consciousness impairment. Despite these negative links, strong positive associations persisted between clinically related symptoms such as vomiting - nausea, weakness- changes in taste perception, balance disorders - weakness, depression - anxiety and fatigue – weakness. Centrality analysis highlighted weakness, appearance issues, and consciousness impairment as the most central symptoms across strength and betweenness, while fatigue also ranked highly in closeness. This indicates that multidimensional decline—physical, psychological, and systemic—was most pronounced in the last month of life. Community detection identified five clusters, summarized in Table 8.

Table 8.

Symptom communities of the first period.

Community	Main theme (Interpretation)	Symptoms included
1	Psychological	Irritability, Depression, Anxiety, Anger
2	Gastrointestinal core	Nausea, Vomiting
3	Neurological (sensory)	Dizziness, Seizures, Blurred vision
4	Neurological (motor/sensory)	Balance disorders, Weakness, Taste changes
5	Broad systemic burden	Fatigue, Appearance issues, Weight loss, Cognitive problems, Itching, Dehydration, Gastrointestinal issues, Pain, Pallor, Consciousness impairment, Pressure ulcers, Oral ulcers/inflammation, Appetite issues, Swallowing difficulties, Sleep problems, Urinary problems, Diarrhea, Respiratory problems, Cough, Constipation, Edema

Accuracy and stability of networks

Tests related to network accuracy were conducted for all six periods. The stability of edge weights was examined using nonparametric bootstrapping. Across periods, the bootstrap confidence intervals (CIs) around the estimated edge weights were generally narrow, indicating that the networks were estimated with good accuracy. Detailed bootstrap plots for each period are provided in the Supplementary Information (Figures S13–S18).

The stability of centrality estimates was assessed using case-dropping subset bootstrapping. As shown in Table 9, the CS coefficient for strength centrality exceeded the recommended threshold of 0.25 in all six periods, ranging from 0.28 (sixth period) to 0.75 (first period), indicating good reliability. In contrast, betweenness centrality showed low stability in several periods, with coefficients falling below 0.25 in the second, fifth, and sixth periods. Closeness centrality was highly unstable across all periods (Full bootstrap plots for centrality stability across all six periods are provided in Supplementary Figures S19–S24).

Table 9.

Correlation stability (CS) coefficients for centrality indices across six periods.

Period	Strength	Betweenness
1 (last month)	0.75	0.59
2 (1–2 months)	0.67	0.12
3 (2–3 months)	0.67	0.28
4 (3–6 months)	0.36	0.28
5 (6–12 months)	0.44	0.20
6 (> 12 months)	0.28	0.12

Closeness centrality showed very low stability across all periods and was therefore excluded from interpretation.

Discussion

This is the first study to examine the symptom network of cancer patients at the EOL longitudinally. Six-time periods before death were considered to analyze the 37 common symptoms in a network of patients with various types of cancer. After conducting network-related tests, the results for each period were extracted. Findings indicated that different symptoms could emerge as central symptoms in each period before death. For example, fatigue, pain, nausea, and anxiety were among the central symptoms in this study. These symptoms have also been highlighted in previous network-based symptomatology studies. For example, the study by De Rooij et al.⁸, which examined a broad range of cancer patients, found that fatigue was the most central symptom, followed by fatigue, pain and nausea, which showed moderate centrality values. Another study that effectively emphasizes the importance of these symptoms is the research by Papachristou et al.¹². They analyzed the symptom network from three dimensions: severity, prevalence, and distress. In the prevalence network, nausea had the highest value across all centralities, while fatigue ranked third in closeness centrality after nausea and appetite loss. Additionally, fatigue was the second most central symptom in the severity and distress networks based on betweenness and closeness centrality measures. Similarly, the study by Rha and Lee⁴⁴ aligned with previous research findings. They assessed the symptoms of 249 cancer patients undergoing chemotherapy. Results demonstrated that fatigue had the highest strength centrality throughout chemotherapy. Also, Anxiety appeared frequently as the second most important symptom in their results. For psychological symptoms, Xu et al.‘s²⁸ study also identified distress symptoms (which overlap with depression and anxiety) and fatigue as having the highest centrality values and, therefore, being the most central symptoms. Their result also showed that nausea and pain had high centrality scores compared to other symptoms.

Alongside the previously mentioned symptoms, we identified weakness, issues with physical appearance, and consciousness impairment as central symptoms in our study. These three symptoms were particularly important for patients due to their persistence. Weakness and issues with physical appearance were repeatedly central in four time periods, while consciousness impairment was repeated as a central symptom in three periods. This highlights how crucial these symptoms are in the final year. However, previous studies didn’t recognize these symptoms as central nodes. One of the main reasons for consciousness impairment is that most of the studies often excluded patients with severe conditions or patients who have consciousness impairments. Excluding patients with severe cognitive conditions causes symptom networks to inherently ignore this symptom. In contrast, existing research on EOL cancer care indicates that consciousness impairment is highly prevalent among patients, especially in their final months of life^15,18,45. A similar challenge exists for body image issues or issues with physical appearance (including pallor). While previous network-based studies excluded severely ill patients, research on body image has demonstrated that this symptom significantly affects patients’ mental health, particularly those in the EOL stage. As cancer progresses, patients often experience profound physical changes, such as weight loss, surgical scars, hair loss due to treatment, and other disease-related physical effects. These changes can deeply impact their self-esteem, emotional well-being, and overall quality of life^46–48. Regarding weakness, most earlier studies did not differentiate it from fatigue or included it in their analyses. However, in this study, we examined weakness as a separate symptom to better understand its role within the symptom network.

Gastrointestinal issues, constipation, urinary problems, and dehydration were also identified as central symptoms in our study in some periods, although they were not found in previous research. However, this could be due to the fact that a significant portion of our sample (over 30%) consisted of patients with gastrointestinal cancers, making these symptoms more prominent in the symptom network. Further network-based studies on patients with advanced cancers at the EOL could help clarify the significance of these symptoms and provide a deeper understanding of our findings.

In addition to identifying central symptoms, this study examined the interconnections between symptoms in patients. The result showed that the edge between vomiting and nausea was one of the most significant and persistent relationships across all six time periods. This finding aligns with previous studies utilizing network analysis. For instance, in a longitudinal study by Kalantari et al.⁹ on patients with breast, gastrointestinal, gynecological, and lung cancers undergoing chemotherapy, it was discovered that the vomiting-nausea edge was not only one of the strongest connections but also a stable symptom pair, observed across three-time points (before the first chemotherapy session, two weeks after the first session, and one week after the second session). Additionally, other studies, including those by Xue et al., Zhou et al., Shim et al., Rha & Lee, and Zheng et al.^{6,28,29,37,44} have also identified the vomiting-nausea edge as one of the most significant connections in cancer patients. It seems that regardless of treatment type, tumor type, or even cancer stage, this edge remains significant for patients. Another significant edge in this study was the relationship between depression and anxiety, which remained stable across five periods before death. Similar results have been reported in studies that used network analysis. In a study by Shim et al.³⁷ on patients with gastric cancer, symptoms were assessed longitudinally at three-time points (before surgery, one week after surgery, and three to six months after surgery). The results emphasized the significant impact of psychological and emotional symptoms on patients’ quality of life. The strongest connections identified in the symptom network included depression-anxiety, distress-sadness, and sadness-anxiety. Similarly, in the study by Rha & Lee⁴⁴, the connection between depression and anxiety was also found to be one of the strongest edges in the network. This study also demonstrated that the connection between these two symptoms remained strong and stable from the pre-chemotherapy phase through the fourth chemotherapy cycle.

Additionally, our study employed the Walktrap algorithm to identify symptom communities across six periods before death. The results revealed five distinct communities throughout the periods: (1) a psychological symptoms community, including depression, anxiety, and irritability; (2) a gastrointestinal symptoms community, consisting of vomiting and nausea; (3) a weakness and physical appearance community, including fatigue, gastrointestinal problems, itching, issues with physical appearance, and pain; (4) a physical decline community, characterized by cognitive problems, consciousness impairment, and swallowing difficulties, along with other physical symptoms; and (5) a balance and vision disturbances community, which included dizziness, balance disorders, and blurred vision.

Studies on symptom clustering in cancer patients have shown limited consistency^8,30,44. This variability arises from multiple factors, including the clustering algorithm, the symptom assessment tool (questionnaire), tumor type and location, patient demographics (age and gender), the timing of assessment (before, during, or after treatment), and even the treatment method itself^49,50. Moreover, there are relatively few studies that have focused on symptom clustering in patients with advanced cancer or those in their final year of life, particularly from a longitudinal perspective, and evidence supporting the stability of such clusters remains limited^49,51. However, some findings from our study are consistent with previous research. For example, nausea and vomiting consistently formed a single cluster, while depression, anxiety, and irritability were typically grouped together. Likewise, a study by Simão et al.⁵¹ on patients with advanced cancer identified a similar nausea-vomiting cluster. The researchers assessed symptoms over 30 days at three-time points (study entry, 15 days after entry, and 30 days after entry) and found that these two symptoms remained present across all assessments. Additionally, a study by Chong et al.⁵² demonstrated that depression and anxiety emerged as a psychological symptom cluster in patients with advanced cancer. Similarly, the systematic review by Dong et al.⁴⁹ examined symptom clusters in patients with advanced cancer and identified four distinct clusters, including the nausea-vomiting and anxiety-depression clusters. Their findings revealed that 44% of the studies reported nausea and vomiting occurring together, which aligns with our results. In our study, these symptoms consistently formed a separate cluster across five time periods without being grouped with any other symptoms. Additionally, Dong et al. found that the anxiety-depression cluster co-occurred with at least one other symptom in 44% of the studies. This finding is also consistent with the psychological cluster in our study, where anxiety and depression were grouped with irritability in every time period. Another symptom cluster identified in our study included fatigue, gastrointestinal problems, itching, and issues with physical appearance, along with diarrhea and appetite problems. Although this exact combination of symptoms has not been specifically reported in previous studies, some of these symptoms have been clustered together in prior research. Many studies have shown that pain, fatigue, and appetite problems often cluster together, sometimes with additional symptoms. For example, in the study by Cheung et al.⁵², fatigue, drowsiness, nausea, loss of appetite, and shortness of breath formed a symptom cluster. Similarly, So et al.⁵⁰ examined symptom clustering in breast cancer patients across three phases: before treatment, during treatment, and after treatment. Prior to treatment, the symptoms that clustered together included pain, fatigue, sleep disturbances, nausea, appetite issues, depression, cognitive difficulties, memory problems, and bowel issues. During treatment, the cluster consisted of pain, fatigue, and sleep disturbances combined with concerns about appearance, difficulties in concentration, bowel problems, cognitive issues, depression, anxiety, and hot flashes. Additionally, the study by de Rooij et al.⁸ identified a symptom cluster that included fatigue, pain, emotional symptoms, loss of appetite, and shortness of breath. These findings suggest that while specific symptom clusters may vary across studies due to different cancer types, treatment phases, and methodologies, certain symptoms, such as fatigue and appetite loss, frequently co-occur.

Many factors influence the structure and composition of symptom clusters (or communities). One key reason for the differences between our findings and those of other studies is the methodological approach used to identify symptom clusters. In most studies, principal component analysis (PCA), cluster analysis, or a combination of both has been employed for symptom clustering. Additionally, many studies utilize the Edmonton Symptom Assessment Scale (ESAS) for symptom assessment, which typically evaluates only nine common symptoms in cancer patients. A systematic review by Dong et al.⁴⁹ on symptom clusters in patients with advanced cancer indicated that PCA was used in 36% of the studies, cluster analysis in 15%, and a combination of both methods in 18%. Together, these two approaches were applied in 69% of the studies analyzed. In contrast, our study evaluated 37 symptoms, providing a more comprehensive assessment. Beyond methodological differences, other variables also influence symptom clustering (communities), as previously discussed. Given these variations, further research employing methodologies similar to ours is needed to validate the identified communities.

Limitations and future direction

While our study provides valuable insights into symptom networks among cancer patients who are at the EOL, several areas warrant further exploration. We focused primarily on the presence or absence of symptoms and their prevalence. However, other dimensions, such as symptom severity and distress, can also provide valuable insights. Future studies should explore these dimensions, particularly when investigating symptoms at the EOL, to enhance our findings and enable comparative analyses. Another limitation of our study was the lack of consideration for covariates, such as treatment methods, types of medications used at the EOL, and whether patients were receiving treatment during their final months. These factors should be taken into account in future research to allow for the estimation of more personalized symptom networks. Additionally, this study examined symptom networks by considering a combination of different cancer types. Finally While this study identifies key elements regarding the prevalence of symptoms, their timing in relation to the period preceding death, and the structure of symptom networks including central symptoms, it does not provide a formal proof of concept. This limitation arises from the observational study design and the constraints of the available data, as discussed above. Future research using experimental or longitudinal designs could provide stronger causal evidence to validate these findings. Also it is recommended that future research analyze the symptom network of each cancer type separately at the EOL to enable more precise analyses and the development of personalized interventions.

Conclusion

This study examined 8,026 cancer patients across 26,318 visits using a longitudinal network analysis of symptoms during the last year of life. By modeling six distinct time periods before death, we identified how central symptoms and their interrelationships evolved over time. Symptoms such as weakness, physical appearance issues, and cognitive impairment consistently emerged as central nodes across multiple periods, while other symptoms—including urinary problems, dehydration, constipation, fatigue, and pain—were period-specific. Persistent associations, particularly between nausea–vomiting and anxiety–depression, further emphasized the importance of monitoring symptom clusters rather than isolated symptoms.

These findings underscore the value of longitudinal network analysis in end-of-life care. Recognizing central symptoms and stable symptom clusters offers actionable targets for timely, period-specific interventions. Clinicians can use these insights to anticipate the evolving symptom burden and deliver more effective, individualized care. Future research should refine these networks by incorporating treatment effects, comorbidities, and patient-reported outcomes to further strengthen symptom management strategies in palliative oncology.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

S.M. conducted the investigation, data curation, formal analysis, and visualization, and wrote the original draft. E.N. supervised the project, contributed to the conceptualization and methodology, performed validation, managed project administration, and reviewed and edited the manuscript. A.M. assisted with supervision, validation, and manuscript review. All authors reviewed and approved the final version of the manuscript.

Data availability

The dataset analyzed in the current study is not publicly available due to healthcare-related restrictions, but can be provided to editors and referees upon request. For data requests, please contact the corresponding author.

Code availability

The code used for data analysis and network modeling in this study is available upon request via the corresponding author.

Declarations

Competing interests

The authors declare no competing interests.