This investigation is a multiphase translational study to develop and temporally validate a nurse-led clinical scoring system for the early identification of sepsis in adult patients with FN. The scoring system integrates routinely available clinical and laboratory data with physiological measurements from an AI-enabled, contactless remote patient monitoring (RPM) system. The RPM’s role differs by phase: in phase I, it functions as an accurate, nurse-independent instrument for capturing single point-in-time vital signs at the onset of FN, which then enter the model as candidate predictors. In phase II, continuous tracking is also used to detect clinically significant RPM alerts within the first 24 hours, evaluated as a secondary outcome across score-defined risk strata rather than as a model predictor.

The study follows a structured 2-phase framework.

This prediction model study and temporal validation protocol is registered with the Clinical Trials Registry-India (registration CTRI/2025/06/088290). The study has received approval from the Institutional Ethics Committee of Amrita School of Medicine (approval ECASM-AIMS-2025‐184). Any substantive modifications to the scoring model, predictor set, or use of AI-enabled RPM data that affect patient safety or study conduct will undergo additional ethics review. This document represents protocol version 2.0, dated September 2026.

The study will be conducted in the hematology and BMT department at Trustwell Hospital, Bengaluru, India, a multispecialty center that manages patients with hematological malignancies and performs BMT. The hospital has an integrated hospital information system that allows comprehensive retrieval of clinical documentation, laboratory results, and treatment history and uses AI-enabled remote patient monitoring technology that continuously measures vital signs at the bedside using contactless sensors placed under the mattress. The AI-enabled RPM system (Dozee NX-Pro) was implemented as part of routine bedside monitoring in the hematology and BMT unit before December 2021. The retrospective cohort start date (December 2021) was deliberately selected to coincide with RPM availability at the study site, ensuring that contactless, RPM-derived vital sign measurements are consistently available across the entire phase I derivation cohort rather than only for a subset of patients. The hematology and BMT units are staffed by trained oncology nurses experienced in managing FN and noticing early sepsis. This clinical environment provides a suitable platform to evaluate a nurse-led sepsis identification scoring system in high-risk immunocompromised patients.

Phase II incorporates a nested interrater reliability and feasibility substudy, as the score is intended for independent bedside application by ward nurses. Three rostered staff nurses of the hematology and BMT unit, selected to span the range of unit experience present on the ward, are trained in a single standardized session using the scoring manual (Multimedia Appendix 1) and 5 written case vignettes. They must meet a predefined competency threshold on the vignettes before participating. For each eligible patient enrolled in phase II, the staff nurse formally assigned to that patient’s care at FN onset calculates the score that feeds the primary phase II diagnostic accuracy analysis. Where at least one additional trained nurse is also on duty and available, that nurse (and a third, where rostering permits) independently calculates the nurse-led score for the same patient within the same 60-minute window, while remaining blinded to the primary rater’s and any other rater’s score sheets; these additional scores are used solely to compute interrater agreement and feasibility for this nested substudy and do not contribute to the primary phase II diagnostic accuracy analysis. Completed sheets are sealed upon submission, and raters may not reconcile or discuss results. As ward rostering will not always place all 3 nurses on the same shift, the number of raters per patient varies between 2 and 3; the agreement statistics used accommodate a variable number of raters per subject, and the achieved number of multiply-rated patients, together with the number rated by 3 and by 2 nurses, will be reported rather than fixed in advance. As multiple-nurse bedside scoring cannot be reconstructed retrospectively, the substudy is restricted to patients enrolled from the date of the protocol amendment onward; we therefore report separately the number of patients participating in the substudy and the number participating in the main validation analysis. Agreement will be quantified using Fleiss κ for the binary low- versus high-risk classification and, for each score item, the intraclass correlation coefficient (2-way random effects and absolute agreement) for the total score, exact percentage agreement, as well as pairwise Cohen κ for each rater dyad, each with 95% CIs and interpreted against conventional agreement bands. Feasibility will be assessed as time to complete the score, the proportion of sheets with any missing item and the missing rate per item, and the proportion of patients scored within the specified window. Item-level Fleiss κ below 0.60 will prompt rewording or reoperationalization of that item in the final scoring manual, reported transparently as a protocol-specified refinement. This substudy is noninterventional, does not alter patient management, and imposes no additional patient burden beyond parameters obtained in routine care; participating nurses provide written consent for the collection of their scoring data.

Vital signs will be monitored continuously using the Dozee NX-Pro remote patient monitoring system, which combines a contactless undermattress sensor (ballistocardiography based) to derive heart rate and respiratory rate and connected devices and plug-in probes to capture blood pressure, oxygen saturation, and body temperature. As RPM was already in routine clinical use at the study site before the start of the retrospective cohort period, the same contactless measurement approach captured physiological predictor variables for all phase I patients, avoiding a mixed-methodology dataset combining RPM-derived and manually charted vitals.

The primary outcome is sepsis in patients with FN, defined using Sepsis-3 as suspected or confirmed infection with an increase of at least two points in the Sequential Organ Failure Assessment (SOFA) score from baseline. Baseline SOFA will be reconstructed from clinical data recorded before or at the onset of FN, with chronic organ dysfunction (eg, preexisting renal or respiratory insufficiency) reflected in the baseline score, so only acute worsening contributes to the required ≥2-point increase.

In both phase I and phase II, the primary outcome is Sepsis-3 sepsis occurring within 14 days of the onset of FN, ensuring a consistent outcome window for model derivation and temporal validation. This 14-day outcome window was adopted via a protocol amendment in September 2026 to align the Phase I and Phase II outcome definitions; the original Phase I outcome definition (Sepsis-3 sepsis at any point during the index admission) and the chronology of this change relative to Phase I model derivation are described in the “Results” section. In phase I, we retrospectively reconstruct SOFA scores from routinely recorded respiratory, cardiovascular, hepatic, renal, hematologic, and neurologic data and compare the calendar date patients first met Sepsis-3 criteria with the date of FN onset to confirm the event falls within the 14-day window. In phase II, we prospectively collect SOFA components from clinical records and monitoring systems within the same window. Patients who meet Sepsis-3 criteria more than 14 days after FN onset are classified as nonevents for the primary outcome in both phases, consistent with the study’s focus on early sepsis detection at the point of greatest clinical actionability.

When full SOFA reconstruction is not possible, we will apply predefined Sepsis-3–aligned surrogate organ dysfunction criteria (Multimedia Appendix 3). In phase I, 2 adjudicators independent of the clinicians who treated the patients will review each retrospective patient with FN independently, using a structured data abstraction sheet and the predefined adjudication algorithm described in Multimedia Appendix 3. Both adjudicators will be blinded to the study’s candidate predictor variables when recorded separately from outcome-relevant data; however, because predictor values and SOFA components are embedded within the same retrospective clinical record, complete blinding to all predictor information cannot be guaranteed for every patient. This represents a limitation of retrospective outcome ascertainment and is discussed later on. A third adjudicator or senior clinician will resolve disagreements between the 2 adjudicators, and interrater agreement for the binary sepsis classification will be assessed using Cohen κ coefficient.

In phase II, 2 assessors independently adjudicate the primary sepsis outcome: a clinician not involved in the patient’s clinical care and a trained research adjudicator, applying the same predefined Sepsis-3–aligned algorithm and structured abstraction sheet used in phase I (Multimedia Appendix 3). Neither adjudicator participates in bedside scoring. Both are blinded to all nurse-led and comparator score values: completed score sheets are stored in a separate, access-restricted file, and adjudicators receive a deidentified clinical extract from which all score sheets, score values, and risk category labels have been removed. Adjudication occurs after the 14-day outcome window has closed, using the locked score record described earlier so that neither assessment can influence the other. Disagreements are resolved by a third senior clinician who is likewise blinded to score values, and interrater agreement for the binary sepsis classification is quantified using Cohen κ with 95% CIs. Phase II therefore conforms to STARD 2015 items 13a and 13b for blinded, independent reference standard assessment. As phase II candidate predictors and SOFA components come from the same clinical record, we cannot entirely exclude residual incorporation bias; we discuss this in the “Limitations” section.

If temporal validation shows poor discrimination or systematic miscalibration, model updating may include recalibrating intercept and slope, re-estimating selected coefficients, or adjusting score cutoffs to improve clinical usefulness. We will clearly report any changes as post hoc updates and consider them preliminary until validated in future studies. We will examine associations between baseline score–defined risk groups and RPM alerts using chi-square tests, and estimate odds ratios with 95% CIs for higher- versus lower-risk groups as exploratory analyses. All statistical analyses will be performed using IBM SPSS (version 29.0) and Python (version 3.13); Python code used for procedures such as bootstrap resampling will be archived to support reproducibility and transparency of the modeling process.

The study protocol received approval from the Institutional Ethics Committee (IEC) of Amrita School of Medicine (registration ECNEW/INST/2023/KL/0379) on April 16, 2025 (approval ECASM-AIMS-2025‐184). Trustwell Hospital, Bengaluru, the designated site for recruitment and data collection, does not maintain its own institutional ethics committee. The IEC formally reviewed and noted the institutional permission letter from Trustwell Hospital as part of its assessment before granting ethical clearance, hence extending ethics oversight to this site. In addition, the principal investigator obtained separate institutional permissions authorizing data access and study conduct from the administration of Trustwell Hospital (dated March 11, 2025) and from the Director of the Bone Marrow Transplantation and Hematology Department at Trustwell Hospital. All prospective participants will provide written informed consent before enrollment in phase II, and the participant information sheet and informed consent form are provided as Multimedia Appendix 4. The study will maintain patient confidentiality by anonymizing all extracted clinical data and restricting access to the study database to authorized research personnel; it will follow the Declaration of Helsinki and institutional research governance guidelines. The AI-enabled system (Dozee NX-Pro) will store remote patient monitoring data on secure institutional computers, with access permissions and audit trails in accordance with hospital policy. The RPM system will be used as a source of contactless point-in-time physiological predictor data in phase I and as a source of continuous monitoring–derived alert data for the phase II secondary outcome; in neither phase will it independently initiate or modify treatment, and clinical responsibility will remain entirely with the treating hematology and bone marrow transplantation teams. We will document and report phase II to the ethics committee in accordance with institutional procedures. As this manuscript reports a study protocol and does not include any person’s identifiable data (including images or videos), specific patient consent for publication is not applicable.

This protocol addresses an important unmet need in early sepsis identification among adults with FN by proposing a context-specific, nurse-led predictive scoring system embedded within hemato-oncology and BMT supportive care. Existing tools provide general risk stratification but have limited sensitivity and discrimination in immunocompromised patients with sepsis, pointing to the need for adapted models in this group [41,42]. The planned score aims to bridge this gap by combining FN–specific clinical and laboratory variables with precisely measured, contactless physiological parameters captured at the onset of FN, so risk prediction better represents the unique pathobiology and rapid clinical course of this population. Conceptually, the score is designed not as a unit-specific instrument but as a structured supportive oncology tool that can guide nurse-led escalation pathways, early sepsis workup, and treatment continuity in high-risk hematology patients.

A key strength of this protocol is its methodological strictness. The nested matched case-control design in phase I allows efficient identification of predictors while controlling for key design variables, such as malignancy type and treatment modality, consistent with recommendations for matched case-control prognostic studies [43,44]. Multivariable conditional logistic regression, followed by transforming regression coefficients into an integer-weighted bedside score, balances statistical strength with bedside usability. Internal validation by bootstrap resampling and a distinct same-center temporal validation cohort align with best practice for prediction-model development and validation, which emphasizes appropriate internal, internal-external, and external validation to avoid overfitting and maintain transportability [45].

The protocol also uses contactless remote patient monitoring to capture accurate point-in-time physiological predictors in phase I, while additionally examining continuous monitoring–derived alerts as a secondary marker of physiological instability in phase II. Continuous assessment has been shown to detect subtle physiologic changes that precede clinically apparent deterioration more effectively than intermittent vital sign checks [46,47]. Remote monitoring equipped with early-warning systems has improved ward safety and the timeliness of escalations in tertiary care hospital wards. By examining first-24-hour RPM alerts alongside clinical and laboratory data, this study may clarify temporal patterns of physiological instability in FN and enhance the granularity of sepsis risk stratification. The nurse-led nature of the scoring system, with a PhD nursing scholar and senior BMT-hematology nurse leading design and application, moves decision support closer to the bedside and is intended to reduce variation in clinical judgment, support earlier recognition of deterioration, and improve communication within multidisciplinary teams.

Risk stratification at the onset of FN directly affects supportive oncology care. A high-risk score can trigger earlier empiric antibiotics and hemodynamic optimization, which may prevent progression to septic shock and avoid unplanned ICU transfer. Conversely, clearer stratification may help clinicians decide between safe ward-based management and timely ICU input. At the same time, identifying truly low-risk patients can support maintenance of planned chemotherapy dose intensity and reduce treatment delays due to sepsis-related instability. This score is intended not only as a bedside assessment tool but also as an enabler of structured, nurse-led supportive oncology care, linking real-time physiological surveillance with escalation pathways that can preserve treatment continuity and reduce acute care burden in high-risk hematology patients [48,49]. This aligns with international developments in proactive supportive oncology and early supportive or palliative care interventions but focuses specifically on the underemphasized domain of sepsis-related risk stratification in FN. As the score provides structured risk categories at the onset of FN, it may also facilitate clearer communication between bedside nurses, oncologists, and intensivists, for example, by prompting risk-stratified triggers for early ICU consultation or high-dependency care [15,49].

Risk stratification tools such as the MASCC risk index and the CISNE have been widely used to identify patients with low- and high-risk FN for outpatient versus inpatient management. Prospective validation of the MASCC risk index showed that a score ≥21 could correctly classify most patients with low-risk FN, with high predictive values for an uncomplicated course [50]. However, subsequent systematic reviews and meta-analyses [22] comparing MASCC and CISNE have highlighted important limitations in sensitivity for predicting serious complications, with performance varying by score and threshold. These results show that oncology-specific parameters may not provide sufficient sensitivity to identify early sepsis risk in immunocompromised patients with FN, particularly in settings where acute clinical deterioration is common.

Physiology-based early-warning scores, such as NEWS and Modified Early Warning Score (MEWS), have also been evaluated for FN. In emergency department cohorts, NEWS and MEWS have shown strong prognostic value for early mortality, sometimes outperforming oncology-specific scores such as MASCC and CISNE. [51]. They can be calculated quickly at presentation using vital signs. At the same time, these scores rely on overt physiological abnormalities and established organ dysfunction and were primarily designed for general acute illness rather than FN-specific sepsis. Their focus on short-term mortality and their dependence on intermittent vital-sign measurements may limit their usefulness for detecting the very earliest stages of FN-related sepsis in continuously monitored hematology patients

Several other FN risk scores and prediction models have been developed for specific purposes, such as predicting first-cycle FN risk in patients receiving myelosuppressive chemotherapy or identifying risk factors for multidrug-resistant bacteremia in FN patients. These models have advanced understanding of FN risk and have informed guideline-based recommendations. Still, they frequently focus on chemotherapy cycle-level risk or specific infection outcomes rather than early sepsis detection at the bedside. In addition, most existing FN scores were developed by previous researchers lacking continuous remote monitoring and do not incorporate granular, time-linked physiological data that could capture subtle deterioration before obvious clinical signs appear [52].

From a nursing perspective, oncology nurses have a central role in early identification and management of neutropenic sepsis and in leading protocol-based pathways. Nurse-initiated neutropenic sepsis pathways have been associated with more timely antibiotic administration and more appropriate initial antibiotic selection, supporting the potential value of structured nurse-led care [53]. 54However, few existing FN risk tools are explicitly designed as nurse-led bedside scores that integrate continuous remote patient monitoring data and FN-specific clinical information into a single, interpretable sepsis risk stratification framework

This protocol builds on existing work by bringing together 3 strands: FN-specific risk stratification, physiology-based early-warning scores, and nurse-led sepsis pathways. It extends FN risk scoring beyond unchanging indices by incorporating contactless, accurately measured physiological data from remote monitoring systems as predictor input in phase I, and by examining continuous monitoring–derived alerts as a secondary marker of physiological instability in phase II, while retaining a regression-based, interpretable model that nurses can apply at the bedside. By focusing on adult hematology and BMT patients with FN and embedding the score within nurse-led supportive care and escalation processes, the study addresses a gap in current practice: tailored FN-specific, nurse-centered sepsis prediction models are not yet available [26].

The researcher should acknowledge several limitations of this study design. The retrospective derivation phase depends on the completeness and quality of routine documentation in the hospital information system and nursing notes. Missing data and variation in how clinical signs or RPM alerts are recorded may affect the availability and accuracy of some candidate predictors and SOFA components. The study is conducted in a single tertiary hematology and bone marrow transplantation unit in an urban private hospital in Bengaluru, which may limit generalizability to centers with different patient profiles, chemotherapy protocols, staffing levels, or access to continuous remote monitoring technology. Future work should examine performance in multicenter cohorts, including hospitals without RPM, and may need to develop a version of the score that relies only on routinely observed vital signs. As phase I physiological predictors are derived from single readings at FN onset rather than continuous trend data, the model does not yet exploit the full continuous monitoring capability of the RPM system; incorporating trend- or variability-based features is identified as a direction for future model refinement, informed by the phase II RPM alert analysis.

As candidate predictors and the reconstructed SOFA-based outcome in phase I come from the same clinical records, and some candidate physiological predictors are conceptually related to SOFA’s cardiovascular and respiratory components or to qSOFA criteria, this overlap may inflate apparent model discrimination. We plan a prespecified sensitivity analysis that refits the model without these overlapping predictors and reports both AUCs side by side to quantify this effect (see “Statistical Analysis” section).

Aligning the phase I and phase II primary outcome to a consistent 14-day window improves comparability of model performance across phases but excludes later-occurring sepsis events from the phase I derivation sample; as bone marrow transplantation admissions often extend beyond 14 days, some patients who develop sepsis later in their admission will not contribute a case event to model development, which may limit the model’s applicability to later-onset sepsis and will be examined further in future work.

Outcome adjudication differs in rigor between the 2 phases. In phase I, sepsis outcomes are adjudicated independently by 2 reviewers who are not involved in the patients’ clinical care, with disagreements resolved by a third reviewer and interrater agreement quantified using Cohen κ; however, because candidate predictor values and SOFA components are both recorded within the same retrospective clinical chart, full blinding of adjudicators to predictor data cannot be guaranteed for every patient, and residual incorporation bias cannot be excluded by the researcher. In phase II, separate individuals perform index test scoring and outcome adjudication, and 2 assessors blinded to all score values adjudicate the primary outcome, in line with STARD 2015 recommendations for independent, blinded reference-standard assessment. Two residual limitations remain. First, because phase II predictor values and SOFA components are both documented in the same clinical record, adjudicators may encounter the underlying clinical variables although score values and risk categories are withheld; residual incorporation bias therefore cannot be entirely excluded, and its expected direction is toward modest overestimation of apparent accuracy. Second, adjudicators work within the same clinical unit and cannot be blinded to whether a participant is enrolled in the study.

Restricting the analysis to the first FN event per patient avoids clustering. It simplifies modeling, but it may reduce direct applicability to patients who experience recurrent FN across multiple chemotherapy cycles. The matched case-control design in phase I means variables used for matching are controlled by design and cannot be fully evaluated as independent predictors in the final model. We now evaluate interrater reliability and bedside feasibility of the nurse-led score prospectively in phase II through a nested substudy in which trained frontline nurses independently score the same patients. This substudy has 5 notable limitations. First, the rater pool comprises 3 nurses from a single unit who received identical training, so agreement may be higher than would be observed after routine, less intensive dissemination to a larger and more heterogeneous nursing workforce. Second, the substudy is designed for descriptive precision rather than hypothesis testing, and with 3 raters the CIs around agreement estimates will remain appreciable. Third, ward rostering means some patients will be scored by 2 rather than 3 nurses, which reduces precision for those patients. Fourth, the substudy applies only to patients enrolled after the protocol amendment because bedside scoring by multiple nurses cannot be reconstructed for patients enrolled earlier in phase II; interrater agreement is therefore estimated in a subset of the validation cohort rather than across all enrolled patients. Finally, the substudy measures whether nurses can apply the score consistently and quickly, but not whether doing so changes escalation behavior, workload, alarm fatigue, or patient outcomes; these remain the objectives of a subsequent implementation study.

Finally, same-center temporal validation in phase II provides an important test of performance over time but does not replace formal external validation. Before the principal investigator can recommend widespread adoption, the score will require multicenter external validation and, if needed, recalibration in diverse oncology settings, including centers with different RPM infrastructure and supportive care models.

This protocol describes a nurse-led clinical prediction model study designed to improve early recognition of sepsis in adults with hematological malignancies who develop FN. By combining FN-specific clinical and laboratory variables with accurately measured, contactless physiological data from remote patient monitoring, and by embedding the score within hemato-oncology and BMT supportive care, the planned model aims to provide more sensitive and timely risk stratification than existing general early-warning or complication scores. The regression-based, interpretable design is intended to make the score usable at the bedside by oncology nurses, supporting earlier escalation, more consistent sepsis workup, and better alignment between continuous physiological surveillance and nurse-led decision-making.

If phase I and phase II confirm good discrimination and calibration and show that the score performs at least as well as MASCC, CISNE, NEWS-2, and qSOFA for FN-related sepsis, the tool could help reduce missed or delayed sepsis recognition, unplanned ICU transfers, and chemotherapy disruption in this high-risk population. However, real-world impact will depend on successful nurse-led implementation, integration with existing sepsis bundles and supportive oncology pathways, and formal multicenter external validation across diverse settings, including centers without advanced remote monitoring capabilities. The longer-term goal is a practical, scalable, nurse-centered bedside score that links continuous monitoring with structured escalation pathways, strengthening supportive care and potentially improving outcomes for hemato-oncology and BMT patients with FN.