Introduction
The goal of palliative care for oncology patients is to relieve physical, psychological, and existential suffering and to enhance quality of life (Drăghici and Bucur Reference Drăghici and Bucur2025). However, symptoms in advanced cancer, such as pain, dyspnea, fatigue, anorexia, cognitive decline, anxiety, and nausea, are highly dynamic and often escalate rapidly due to disease progression, treatment toxicity, systemic inflammation, metabolic disturbances, or psychological stressors (Barina et al. Reference Barina, Newcomer and Reddy2020; Jerop Reference Jerop2025).
In most healthcare settings, symptom management is reactive; that is, clinical action is only taken when patients report severe discomfort or when objective changes are clinically apparent (Weiss et al. Reference Weiss, Franck, Leutwyler, Smith and Liehr2023; Clark and Kleinig Reference Clark and Kleinig2024). Such delayed responses can cause unnecessary suffering, unplanned hospital admissions, use of emergency departments, poor quality of life, and distress for carers (Bayuo et al. Reference Bayuo, Agbeko and Acheampong2022; Cotogni and De Luca Reference Cotogni and De Luca2022).
Emerging advances in clinical prediction provide opportunities to move from reactive to proactive symptom management (Donnelly et al. Reference Donnelly, Neff and Sedlack2025; Vaiphei Reference Vaiphei2025). Laboratory biomarkers including C-reactive protein (CRP), neutrophil–lymphocyte ratio (NLR), and abnormal levels of albumin, ferritin, and electrolytes have demonstrated strong associations with systemic inflammation, metabolic instability, cachexia, cognitive decline, and overall clinical deterioration (Padilla-Raygoza Reference Padilla-Raygoza2024; Li et al. Reference Li, Deng and Xie2025). These biomarkers often change before the escalation of subjective symptoms and are therefore useful as early-warning indicators (Rettori Reference Rettori2024 ). At the same time, predictive analytics are leveraging machine-learning models and electronic health record (EHR) data to study trends and find patterns that can predict symptom trajectories more precisely than typical clinical assessment alone (Parikh et al. Reference Parikh, Hasler and Zhang2022; Swanson et al. Reference Swanson, Wu and Zhang2023).
Nurses are key to the successful application of predictive symptom management because they are continuously observing patients, identifying subtle clinical changes, providing context for laboratory abnormalities, communicating with interdisciplinary teams, and initiating timely interventions (Alotaibi et al. Reference Alotaibi, Alharbi and Al-Harbi2023). They use predictive signals and clinical judgment to gain a sophisticated understanding of patient behavior, emotional state, and physiological trends (Hijazi et al. Reference Hijazi, Alghamdi and Alqarni2024). While individual studies have investigated predictive biomarkers, nurse-led interventions, and predictive algorithms, there has been no comprehensive synthesis of how these elements combine to inform nurse-led predictive symptom management in oncology palliative care. This systematic review fills this gap by synthesizing the evidence on the joint prediction of symptom escalation and patient outcome improvement based on laboratory indicators and nursing judgment.
Methods
Review design
This systematic review was performed according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines. This review sought to synthesize empirical studies that examined predictive symptom management delivered by or integrated with nurses in oncology palliative care. The overall objective was to assess the contribution of laboratory biomarkers and nursing clinical judgment in the early detection of symptom deterioration, the predictive accuracy of different indicators and models, and the impact of predictive approaches on patient outcomes such as symptom burden and quality of life. The review protocol was registered prospectively with the International Prospective Register of Systematic Reviews (PROSPERO; Registration No. CRD420261432500).
Search strategy
To maximize coverage of research on predictive symptom management in oncology palliative care, we developed a broad search strategy. Searches were performed in PubMed, Scopus, CINAHL, Embase, and the Cochrane Library from inception to November 2023. The search strategy used combinations of controlled vocabulary and free-text terms related to nursing, palliative care, oncology, predictive models, biomarkers, symptom escalation, and patient outcomes. Boolean operators were used to refine and expand searches where appropriate. The year of publication was not restricted. We also manually searched the reference lists of all included studies for other eligible publications not identified through database searching.
Eligibility criteria
Eligibility criteria were developed for the inclusion of studies directly relevant to the review’s objectives. Studies were eligible if they were conducted with adult oncology patients receiving palliative or supportive care and if they studied nurse-led or nurse-integrated symptom monitoring or predictive interventions. Studies were required to include laboratory biomarkers, risk-scoring tools, predictive models, or other quantifiable indicators of symptom escalation and report outcomes such as symptom burden, predictive accuracy, clinical deterioration, or quality-of-life (QOL) measures. Eligible designs were randomized controlled trials, cohort studies, quasi-experimental studies, mixed-methods research, implementation studies, and predictive modeling papers. We excluded studies with pediatric populations, studies without empirical data, studies only predicting survival with no symptom-related outcomes, or studies only employing physician-led assessments in their predictive processes.
Study selection procedure
The selection of studies was carried out in several steps according to the PRISMA 2020 recommendations (Figure 1). All records were imported into a reference management system, and duplicates were removed. Two reviewers independently screened titles and abstracts of potentially eligible studies. The remaining studies were then obtained in full text and were evaluated against the inclusion and exclusion criteria. Throughout the process, differences were resolved by discussion and, where required, by consensus with a third reviewer. Transparency and reproducibility were assured by documenting the reasons for exclusion.
PRISMA flow diagram summarizing the study selection process.
Figure 1 Long description
The study selection process via databases and registers begins with ′Identification′ where records are identified from databases (n = 1,482) and registers (n = 0). Records removed before screening include duplicate records (n = 346), those marked ineligible by automation tools (n = 8) and others removed for various reasons (n = 16). This results in 1,136 records screened. In the ′Screening′ phase, 1,008 records are excluded. Reports sought for retrieval number 128, with none not retrieved. Reports assessed for eligibility remain at 128. The ′Included′ phase shows 14 studies included in the review, with 14 reports of included studies. Reasons for exclusion are listed: not nursing-focused (n = 42), not predictive symptom management (n = 38), no laboratory biomarkers included (n = 21) and not oncology palliative care population (n = 13).
Quality appraisal
We used appropriate and validated tools based on the study design to perform a structured quality assessment of each included study to assess the methodological rigor and risk of bias. Randomized controlled trials were assessed using the Cochrane Risk of Bias 2 framework. Observational studies and quasi-experimental research were assessed for confounding, selection bias, measurement validity, and reporting of outcomes using the ROBINS-I tool. The quality of the predictive modeling studies was evaluated using the PROBAST tool, including predictors, outcomes, sample size, overfitting, and validation of the model. For mixed methods and qualitative studies, components were assessed using the CASP qualitative checklist. Quality appraisal results were used in the synthesis to provide context to the findings.
Data extraction and synthesis
A standardized data extraction form was used to extract key details from each included study, including authorship, year, country, study design, sample characteristics, biomarkers or predictive indicators used, nursing roles, predictive approach, patient outcomes, and primary findings. The extracted data were reviewed for completeness and accuracy. Given the heterogeneity in study designs, biomarkers assessed, predictive tools used, and outcomes measured, we chose a narrative synthesis approach. Results were thematically categorized into biomarker-based prediction, nursing-led interpretation, predictive algorithms, and QOL effects.
Results
Characteristics of included studies
Fourteen studies fulfilled all inclusion criteria (Table 1). They included randomized controlled trials, prospective and retrospective cohort studies, mixed-methods studies, telehealth interventions, and predictive-modeling studies. The sample sizes or EHR records in the sample ranged from 60 to >3,400 participants. Laboratory biomarkers, including CRP, NLR, albumin, ferritin, and electrolytes, have been consistently observed across studies to have early predictive value for symptom escalation. Nurses played a key part in translating laboratory abnormalities and correlating them with clinical clues and triggering timely interventions. Eight studies examined predictive algorithms, including random forest, XGBoost, and neural networks, with high accuracy but limited clinical application.
Characteristics of included studies (N = 14)
Table 1 Long description
Columns list Authors year, Country, Study Design, Sample or Setting, Biomarkers or Predictive Indicators, Nursing Role, Key Outcomes, and Relevance to Review. Fourteen studies span the USA, South Korea, Japan, Australia, Norway, China, and Saudi Arabia, with designs including randomized controlled trials, pragmatic randomized controlled trial, cohorts, cross sectional studies, mixed methods, and machine learning model studies. Sample sizes range from 150 oncology nurses to 12,000 palliative episodes, with settings including outpatient oncology, telephonic programs, palliative care patients, cancer registry cohorts, and nurse led palliative services. Predictive indicators include symptom scores and functional decline indicators, longitudinal symptom trajectories and risk flags, risk stratification and psychosocial screening indicators, and lab based markers such as C reactive protein, albumin, neutrophil to lymphocyte ratio, blood urea nitrogen, sodium, electrolyte disturbances, platelet related ratios, systemic immune inflammation index, and neutrophil percentage to albumin ratio. Nursing roles include nurse led primary palliative care delivery, telephonic nurse led palliative care, nurse phone follow ups with advance care planning facilitation, structured symptom scoring, delirium and neuropsychiatric screening, monitoring clinical deterioration with escalation, and assessments feeding machine learning models and registry datasets. Reported outcomes include improved symptom burden, communication, and quality of life; reduced acute care use and emergency visits; increased hospice use and advance care planning documentation; biomarker associations with psychological distress, delirium, agitation, cognitive decline, complications, depression, fatigue, survival, rapid decline, and symptom burden; and predictive models for delirium, palliative phase transitions, and referral needs. One study focuses on oncology nurses in a tertiary cancer center and reports training gaps affecting confidence in symptom evaluation and palliative care needs detection.
Discussion
Biomarker-based prediction: Strengths, nuances, and clinical implications
The consistent results across studies provide strong support for the use of laboratory biomarkers as early indicators of deterioration in palliative oncology care. CRP and NLR were often reported as reliable markers of inflammation and systemic stress, often preceding clinical deterioration. For instance, a high ratio of CRP to albumin (CAR) was independently correlated with short-term mortality in patients with terminal cancer, especially during the first 30 days of palliative admission (Ju and Ma Reference Ju and Ma2020). This time lead suggests that elevated CAR may be a marker of decline to come, before the rise of subjective symptoms.
In line with this, a large prospective cohort study in Brazil showed that CRP, albumin, CAR, and even the modified Glasgow Prognostic Score had good discriminatory capacity (C-statistic > 0.70) for 90-day mortality in patients with advanced cancer. The best predictive value was shown by CAR (Silva et al. Reference Silva, Oliveira and Wiegert2024). This is consistent with earlier work from Fudan University that CRP/albumin ratio 1.31 was a significant predictor of worse survival in a palliative-care cancer population (HR 2.33) (Zhang et al. Reference Zhang, Zhang and Li2019).
Moreover, the CAR-mortality association appears to be consistent across settings: for non-small-cell lung cancer patients admitted for palliative care, Karahan and Yalçin (Reference Karahan and Yalçin2020) showed that baseline CAR was significantly higher in non-survivors, suggesting potential utility in disease-specific prognostication. Their retrospective analysis indicated that CAR may be a low-cost and readily available marker that can be used to stratify risk, which is highly attractive for routine nursing-led monitoring.
In addition to CAR, other combinations of biomarkers were also identified in the reviewed studies, including NLR, PLR, SII (Bu et al. Reference Bu, Liu and Kong2025), and NPAR (Wang et al. Reference Wang, Liu and Liu2024) as predictors of physical complications, psychological distress, and fatigue. These multi-parameter ratios may permit a richer, more nuanced prediction algorithm that may reflect both inflammatory and nutrition/metabolic status.
However, while the evidence is compelling, some caveats do arise. McFarland et al. (Reference McFarland, Applebaum and Bengtsen2022) found that albumin and NLR abnormalities correlated with psychological symptoms such as depression and anxiety. The overlap suggests that biomarker changes are not always interpretable as purely physiological deterioration, but may reflect psychophysiological distress, blurring the signal for “hard” clinical decompensation. Thus, interpretation of biomarker changes without simultaneous clinical assessment is prone to over- or under-interpretation.
Nursing-led interpretation and clinical decision-making: Core to predictive care
Nurses are not just implementers but also interpreters and decision-makers for predictive symptom management. The studies consistently show their central role in our review. For instance, in the cluster RCT “CONNECT” intervention (like Reinke’s model), oncology nurses conducted regular visits, identified symptoms using ESAS, and collaborated on shared-care plans (SCPs) with patients (Symptom Burden & Shared Care Planning sub-study) (Mitchell et al. Reference Mitchell, Althouse and Feldman2023). These SCPs targeted the most troublesome symptoms, and nurses frequently used non-pharmacological interventions (70% of plans), indicating a patient-centered, nuanced approach rather than a purely clinical escalation model. This shows an example of how nurses utilize patient-reported experiences along with clinical assessment to guide care.
This was complemented by Reinke et al. (Reference Reinke, Sullivan and Slatore2022), who showed in their RCT that nurse-led primary palliative care (via telephone) maintained or improved symptom burden and quality of life over time (although their 3-month outcomes were not statistically different to a large extent). This implies that nurses are able to provide meaningful and sustained monitoring, albeit remotely, which is essential for predictive care paradigms.
Kim et al. (Reference Kim, Lee and Woo2024) built an ML model for predicting delirium risk in palliative-care patients in a large international study with 36 clinical and lab predictors (including CRP, BUN, sodium, and NLR). Nurses played a key role in gathering structured assessments that fed into the model, bridging the gap between raw data and predictive insight. Similarly, Kawashima et al. (Reference Kawashima, Furukawa and Imaizumi2024) developed predictive models (using CRP, albumin, ECOG, and symptom scores) to predict palliative-care referral needs. They relied on symptom assessments scored by nurses and stressed the necessity of nurse-led data collection and interpretation.
On the Saudi side, Almutairi et al. (Reference Almutairi, Alsubhi and Ghurab2023) found that confidence and perceived competence of oncology nurses in assessing symptoms and palliative care needs were important for early detection of deterioration. Their survey points to a real-world barrier: if nurses do not have enough training and confidence, they may underuse or misinterpret early clinical signs and so fail to realize the full promise of predictive approaches, though not predictive in the biological sense.
Predictive algorithms: Promise, accuracy, and implementation challenges
The use of ML and predictive models in our review suggests a clear move from reactive to proactive palliative care. Guo et al. (Reference Guo, Dai and Jiang2025) developed a model to predict transitions between palliative phases using data from the Palliative Care Outcomes Collaboration (PCOC) registry, which includes nurse-collected functional, symptom, and lab measures. Their model showed promising predictive capacity, indicating that real-time decision support dashboards could be developed to help nurses and clinicians act on early signs of deterioration.
But accuracy is not enough. A major hurdle remains implementation. The best-performing models still face “last mile” challenges, such as how to integrate into EHRs, how to deliver alerts without causing alarm fatigue, and how to gain the trust of nurses so that they act on the alerts. The relative scarcity of studies showing real-world use of predictive models (rather than proof-of-concept) in our review echoes broader concerns from implementation science.
Much of this is similar to observations in other settings: in early palliative care trials, algorithmic or risk-score alerts may not significantly change outcomes if nurses or clinicians do not act on them (or if the thresholds are not actionable). For example, although not in oncology, predictive-mortality models in the hospital setting (Blanes-Selva et al. Reference Blanes-Selva, Ruiz-García and Tortajada2021) achieved AUCs of ∼0.86–0.91 using machine learning, but were not translated into actual palliative referral or intervention.
One big challenge is interpretability. Nurses and clinicians may resist “black-box” models that lack transparent logic or clear protocols for what to do when an alert fires. Future development should therefore not only optimize predictive accuracy, but also embed human-centered design: involve nurses in threshold-setting; develop training; co-create dashboard interfaces; integrate into nursing workflows; and continually evaluate impact. Another challenge is variability in data quality: predictive models require consistent, accurate inputs (e.g., lab draws and timely symptom scoring). In environments with lower frequency of lab access, limited resources, or staffing constraints, such models may perform less reliably or produce missing data, reducing their utility.
QOL and clinical impact of predictive nursing interventions
Another strong argument is that predictive nursing improves patient quality of life. The included RCTs and cohort studies show strong evidence that proactive nursing reduces symptom burden, prevents emergency visits, and increases patient and carer satisfaction (Reinke et al. Reference Reinke, Sullivan and Slatore2022). Although their nurse-led telephone-based intervention did not result in a statistically significant improvement in QOL at 3 months (FACT-L TOI), the study showed feasibility and acceptability of nurse-led primary palliative care for a wide range of patients (including earlier-stage lung cancer).
This large pragmatic RCT (Grudzen et al. Reference Grudzen, Flannery and Van Allen2025) demonstrated that telephonic nurse-led palliative care reduced acute care use and improved patient-reported outcomes, a direct illustration of a proactive, remote, nurse-driven model improving care experience. Liddicoat-Yamarik et al. (Reference Yamarik, Chiu and Flannery2023) found their 6-month nurse-led program increased ACP documentation and hospice engagement, suggesting that early and ongoing nurse follow-up can change trajectories toward anticipatory patient-preferred care. There is also an indirect benefit through predictive biomarkers and algorithms. If deterioration is detected earlier by nurses (e.g., high CAR, NLR, or algorithm alerts), they can intervene (change meds, counsel, and escalate care) before the condition worsens. This kind of pre-emptive care is more likely to stabilize patients, maximize comfort, and prevent unnecessary hospitalizations; all of which improve QOL.
However, note that a number of symptom domains may not be addressed equally: in the SCP analysis of the trial, although symptoms like tiredness, poor appetite, and low well-being were frequent, only a modest proportion of moderately severe symptoms were actively addressed (e.g., only 35% of visits for tiredness via SCPs) (Mitchell et al. Reference Mitchell, Althouse and Feldman2023)
Systemic barriers, contextual challenges, and future directions
While the review highlights strong promise, several systemic and contextual barriers must be addressed for predictive nursing models to succeed broadly:
1. Training and competency: Almutairi et al. (Reference Almutairi, Alsubhi and Ghurab2023) found that oncology nurses in Saudi Arabia had lower confidence in terms of palliative assessment and predictive reasoning and identified substantial educational gaps. Such gaps could prevent them from converting biomarker or algorithmic alerts into timely, effective interventions. Even sophisticated tools can perform poorly if they are not specifically trained to predict, assess risk, and prioritize symptoms.
2. Standardization and thresholds: There is substantial heterogeneity in biomarker thresholds (e.g., CAR cutoffs and CRP levels) across studies. No standardized, evidence-based protocol exists for interpretation of these labs for forecasting nursing workflows. Consensus guidelines or institution-specific thresholds (e.g., by local validation studies) will be essential.
3. Infrastructure and access: The predictive model requires consistent, frequent lab testing and symptom assessments. In situations where resources are scarce (or where access to labs is limited), the model may fail because data are missing or delayed. For instance, strategies for implementation will have to take into account local infrastructure and capacity, such as point-of-care testing, optimizing the frequency of lab orders, or using telehealth for reporting symptoms.
4. Workflow integration and human factors: A design challenge for integrating predictive models into real-world nursing workflows is that alerts need to be meaningful, not overwhelming, action protocols need to be clear, and trust needs to be built. Co-design of dashboards with nurses, pilot of alert thresholds, and measurement of alert fatigue will be key steps. In addition, strategies for change management will be required for adoption and sustainability.
5. Ethical and communication problems: Ethical and communication challenges of deterioration prediction for nurses – how do they communicate risk prediction to patients and families? What are the steps to take, escalate, or recommend when an algorithm introduces a high risk? This will require development of communication training and ethical frameworks for predictive palliative care.
6. Research and validation needs: Future research should prioritize:
o Prospective validation of biomarker-based and ML models in diverse patient populations (including underrepresented geographies such as the Middle East).
o Implementation trials of decision-support tools embedded in clinical workflows.
o Longitudinal studies to assess how predictive interventions impact emergency utilization, hospitalizations, survival, and QOL.
o Cost-effectiveness analyses to justify investments in labs, predictive infrastructure, and nurse training.
Strengths and limitations of the review
Strengths:
• To our knowledge, this review is the first to combine nurse-led, predictive symptom management interventions with laboratory biomarkers and predictive algorithms in oncology palliative care, rarely discussed in conjunction.
• Multiple settings (USA, Asia, Saudi Arabia, and Australia) increase the generalizability of the results and suggest cross-cultural considerations.
• By combining clinical outcome studies and modeling research, we get a complete picture – not only of whether models work but also of how nurses can use them in practice.
Limitations:
• Most included studies of predictive algorithms remain proof-of-concept, with few studies evaluating real-world clinical integration, thus limiting conclusions about feasibility.
• Biomarker prognostic studies have different cut-offs and may be subject to selection bias (patients who have frequent lab draws may not be representative of all palliative patients).
• Publication bias may exist because negative or null implementations of nurse-led predictive care may not be published.
• The studies were heterogeneous in design (RCTs, retrospective cohorts, and ML models), so quantitative meta-analysis was not appropriate.
Conclusion (implications for practice and future research)
To summarize, the converging evidence of your 14 included studies supports a paradigm shift in oncology palliative care from reactive symptom response to proactive, nurse-led, biomarker-informed predictive management. Biomarkers like CRP, CAR, NLR, and multi-parameter ratios offer early warning signs of decline. Nurses interpreting these alongside clinical cues and using predictive models can intervene early, stabilize symptoms, and improve QOL. For this to become a reality, healthcare systems need to invest in training nurses, integrate predictive algorithms into workflows, standardize biomarker thresholds, and conduct real-world implementation research, especially in areas like Saudi Arabia where nursing predictive capacity is still in its infancy. This approach has tremendous potential: it enables nurses, provides better patient support, and may reduce preventable hospitalizations, resulting in palliative care that’s more proactive, person-centered, and respectful.
Acknowledgments
The authors would like to acknowledge all researchers whose work has contributed to the evidence synthesized in this systematic review. We thank the experts in oncology and palliative care who influenced the thinking that led to this study. Special thanks are due to the librarians and research support teams who helped to fine-tune the search strategy and to ensure that all databases were comprehensively searched. And finally, we honor the enduring commitment of nurses and healthcare professionals whose dedication to patient-centered care continues to inspire progress toward predictive symptom management.
Author contributions
A.M.I.: Conceptualization, methodology, literature search, data curation, formal analysis, writing – original draft, and validation. D.E.F.Z.: Conceptualization, methodology, formal analysis, writing – review and editing, critical revision for important intellectual content, and validation. H.A.: Conceptualization, methodology, formal analysis, writing – review and editing, critical revision for important intellectual content, and validation. All authors contributed substantially to the conception and design of the study, participated in the interpretation of the literature and findings, critically revised the manuscript for important intellectual content, approved the final version for submission, and agree to be accountable for all aspects of the work.
Funding
The authors extend their appreciation to Prince Sattam bin Abdulaziz University for funding this research work through the project number (PSAU/2025/03/38525).
Competing interests
The authors declare that they have no competing interests.
Open access
