In-hospital Mortality Prognostication for Cancer Patients with Febrile Neutropenia: A Single Center Observational Study

doi:10.21203/rs.3.rs-4730716/v1

Download PDF

Article

In-hospital Mortality Prognostication for Cancer Patients with Febrile Neutropenia: A Single Center Observational Study

https://doi.org/10.21203/rs.3.rs-4730716/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Febrile neutropenia (FN) in cancer patients undergoing chemotherapy can result in life-threatening outcomes. Hence, an evaluation of associated risk factors can enable clinical surveillance as well as inform prophylactic measures. In this retrospective cohort study, we report a mortality prognostication model for chemotherapy-treated cancer patients upon a neutropenic episode.Clinical and diagnostic data of 137 febrile neutropenia patients (>18 years) was collected from a cancer hospital, with the primary endpoint of post-hospital admission mortality within 30 days. The data was integratively analyzed and machine learning techniques were applied to develop the predictive model which was then internally cross validated. Towards enabling personalized risk assessment, a nomogram was constructed and validated. Chemotherapy-treated cancer patients undergoing a neutropenic episode exhibited an overall mortality rate of 17.36%. Multivariate logistic analysis elucidated that shock, pneumonia, carboplatin, doxorubicin, antifungal and antiviral prophylaxis, and hemoglobin correctly classified cases with an overall accuracy of 92% and discriminated mortality with a specificity of 76%. Antiviral (odds ratio (OR): 0.669, p = 0.689), and antifungal prophylaxis (OR: 0.619, p = 0.5) demonstrated a protective effect. The receiver operating characteristic (ROC) curve of the nomogram exhibited an area under the curve of 0.878 (95% CI 0.778 - 0.977), Hosmer–Lemeshow test p-value = 0.635, and a high net benefit in the clinical decision curve. The proposed model offers insights into the role of clinical predictors as well as treatment characteristics that can ameliorate mortality risk in cancer patients with FN. The study highlights bacteremia-related surveillance, along with thrombocytopenia, linked to carboplatin, for reducing individualized mortality risk along with improved monitoring and informed treatment strategies.

Health sciences/Health care/Public health/Epidemiology

Health sciences/Risk factors

Health sciences/Oncology/Cancer/Cancer epidemiology

Health sciences/Medical research

Febrile neutropenia

Prognostication

Chemotherapy

Mortality

Cancer

Febrile neutropenia (FN) is a significant complication in cancer patients undergoing chemotherapy, impacting 10%-50% of patients with solid tumors and rising to 80% in hematological malignancies^1–3. FN entails considerable risks, including hypotension, acute kidney injury, and respiratory and heart failure, that occur in 25%-30% of cases ⁴. Beyond these primary complications, FN increases susceptibility to invasive infections that can swiftly progress to conditions like septic shock and mortality^5,6 which contribute to an overall mortality rate between 2–21% ^7,8.

Risk factors for FN-related complications of FN encompass age, cancer type, comorbidities, delayed antibiotics, and laboratory or vital sign abnormalities ^9–11. The Talcott, Multinational Association for Supportive Care in Cancer (MASCC), and Clinical Index of Stable Febrile Neutropenia (CISNE) risk scores have been developed using different combinations of these factors, towards enhancing FN management ^10,11. While Kuderer et al ¹² identified an increased risk of hospitalization and post-FN mortality, Nordvig et al ¹³ explored the heightened long-term risk of infections in surviving patients who experienced FN during chemotherapy. Overall, the impact of FN on mortality remains underexplored and hinges on various factors contributive towards noninfectious-related mortality including organ damage, chemo-related dose modifications and non-compliance, patient characteristics, and type of malignancy ^12,14. This underscores the need for effective monitoring and management of cancer patients, not only during chemotherapy, but also for mitigating the long-term threat of severe infections associated with FN.

FN is underpinned by a complex interplay of multiple risk factors and can lead to adverse clinical outcomes. This necessitates the development of integrative clinical decision-making models for its effective and timely management¹⁵. Literature furnishes evidence on the association between likelihood of developing FN and the duration and severity of neutropenia in both acute leukemia patients¹ and in those with solid tumors¹⁶. The risk of febrile neutropenia (FN) increases with both the severity and duration of neutropenia and is most frequently observed early in the course of chemotherapy^17–19. Identifying risk factors associated with increased morbidity and mortality in relation to FN²⁰ is crucial for evidence-based clinical monitoring and prophylactic interventions. Towards this goal, we conducted a multivariate analysis to identify patient-centric risk factors associated with FN mortality and developed a prognostic model for a cohort of FN patients with different cancer types, at a single cancer center. The proposed prognostication model can be used to assess mortality risk after the initiation of a neutropenic episode amongst cancer patients. An improved quantitation of individualized mortality risk would facilitate a more informed strategy for patient monitoring and treatment. Moreover, prognostic modelling could provide substantial evidence towards improving health care systems and advancing medical research especially in countries with limited healthcare infrastructure.

Setting and study population

This retrospective cohort study focused on patients diagnosed with FN at Shaukat Khanum Memorial Cancer Hospital and Research Centre (SKMCH&RC), from Aug 2014 to April 2018 (IRB EX-01-02-22-01). The study had the IRB requirement of written informed consent waived in accordance with documentation of informed consent (45 CFR 46.117) ²¹, as it posed “no more than minimal risks to participants”. In conformance to the definition^22,23, the inclusion criteria was (1) diagnosis of malignancy; (2) with a low neutrophil count (< 1,500/µL); (3) fever defined as an oral temperature (≥ 38.3°C / 101°F); (4) a history of chemotherapy administered within 15 days before their hospitalization and, (5) over the age of 18 years.

The primary objective of this pilot study was to evaluate the risk factors of, and estimate the risk associated with mortality after the onset of FN episode amongst cancer patients. To ensure generalizability, clinical relevance, and practical applicability, SKMCH&RC Hospital Information System (HIS) was mined to identify FN patients and associated datasets, towards identification of multiple patient-centric risk factors associated with FN mortality and onward development of prognostic models. Initially, 1000 patients experiencing neutropenic sepsis were sampled from HIS, from which 137 patients which satisfied the inclusion criteria, were selected (see Supplementary Figure S1).

Based on the power analysis conducted using the compromise method²⁴ results indicated that with a sample size of 137 and a balanced alpha and beta levels of 0.05 ensured sufficient power of 85.7% to detect an effect size of approximately 0.5 in the logistic regression model.

Data collection and Prognostic factors

The study involved a data collection exercise which encompassed patient-related information such as anthropometrics & demographics, oncology history, chemotherapy and prophylactic treatment, microbiology, radiology, pathology & hematology reports. For a detailed overview of the patient oriented clinical features (see Supplementary Table S1). Patient data was systematically curated from the HIS by a consortium of oncologists, radiologists, and physicians, who had access to both structured and unstructured information within the HIS. Each patient was identified by their unique medical record number (MRN). Admission dates were confirmed via the admission history notes. Upon hospital admission after a febrile neutropenic episode, the following baseline clinical data were collected: age; gender; demographics; BMI, comorbidities; type of malignancy; cancer stage; number of chemo cycles; presence or absence of metastasis, chemotherapeutic and antibacterial regimen; use of granulocyte colony-stimulating factor (G-CSF); use of antibacterial, antifungal, antiamoebic, antiviral and antifungal prophylaxis 2 weeks prior to FN episode. Diagnostic data was accessed from the designated report section, emphasizing both pathology and radiology reports. Cancer type information was gathered from biopsy reports and consultant notes, which provided essential histopathological details and clinical insights into cancer classification and staging. Treatment data was collected by an oncologist through examining medication histories and chemotherapy administration notes, detailing the types of chemotherapy drugs used, dosages, schedules, and any recorded side effects. Microbiological data was extracted from the microbiology section within the main pathology reports. It included information on detected pathogens, sample sources, and antibiotic resistance profiles. FN episodes were categorized into three groups based on infection type: (1) clinically documented infection (CDI) defined as fever and local inflammation without confirmation of pathogenic organisms, (2) microbiologically documented infection (MDI) defined as fever with detected infectious organisms from cultures (blood, sputum, urine, stool, throat swab), and (3) fever of unknown origin (FUO) defined as isolated fever without signs or symptoms suggestive of clinical infection and microbial documentation. In hospital case mortality outcome was labelled upon patient demise within 30 days after the diagnosis of FN. Finally, the resulting dataset was audited by a team of oncologists, radiologists and physicians.

Statistical analysis

Quantitative parameters were reported as median and interquartile range (25th − 75th percentile) after being tested for normality whereas categorical variables were expressed as frequency and percentages. Continuous variables were compared using the Mann-Whitney U test while qualitative variables employed Chi-square and Fisher's exact test. Linear discriminant, logistic regression, and random forest classification model with 800 trees and balanced sampling was trained on the neutropenic data set (samples provided in Supplementary Figure S10) and classification accuracies were compared. This dataset didn’t account for any missing values. Interaction terms, collinearity, goodness-of-fit, the presence of confounders or suppressors, and the models’ assumption were tested for the final model. Statistical analyses were performed using IBM SPSS, version 26.0 (IBM Corp., Armonk, N.Y., USA) and R (version 4.3.1). Power analysis was performed using G*Power version 3.1. A two-tailed p value less than 0.05 was considered as statistical significance.

Logistic Regression

Binary logistic regression is the generalization of linear regression to fit an additive model. Regularization was used to limit model complexity and control overfitting²⁵. The simplified final model was internally validated using ten-fold cross-validation and bootstrapping to assess the stability and reliability of the estimates. We constructed a nomogram²⁶ based on the final fitted logistic regression model, to translate each effect within the model onto a 0 to 100 scale, maintaining proportionality with the log odds. The points were summed across predictors to yield Total Points, which subsequently transformed into a linear predictor and provided predicted probabilities²⁷. The model fit employed a binomial distribution with a logit link function. Next, the distribution of potential predictors in the model, along with their aggregate regression score, was superimposed on the nomogram scales. The accuracy of the model in predicting absolute risk was assessed using the Hosmer-Lemeshow test and the calibration graphical method with a p-value < .05 indicating poor calibration. Decision curve analysis (DCA) was employed to evaluate the clinical practicality of the nomogram model.

Interpretability

Machine learning (ML) methods are inherently variable in their interpretability. Linear models establish explicit connections between clinical inputs and predictions. In contrast, ensemble methods, like random forests aggregate multiple individual models, thus limiting their interpretability. Notably, models lacking clinical interpretability pose considerable challenges in justifying predictions and evaluating validity, thereby impeding clinical translation²⁸. Model sparsity - another aspect of interpretability, explains the number of features used for generating a prediction. For that, we split input features into 3 main branches (see Supplementary Figure S2 & S3) for better feature extraction.

1. Patient Characteristics

A sample of 1000 hospitalized cancer patients over the period of 44 months (Aug 2014 to April 2018), recognized for neutropenia sepsis, were collected. Of these patients, 137 (13.7%) were categorized to develop febrile neutropenic sepsis using neutrophil count (< 1,500/µL), fever (≥ 38.3°C / 101°F) thresholds^22,23 (Supplementary Figure S1). The neutropenic cohort reported a mortality rate of 18.24% amongst patients, who suffered from Grade 3 or 4 neutropenic sepsis. Males exhibited a higher rate of deaths as compared to females (64% vs 36%). There was no contrasting difference seen amongst alive and deceased patients for age, BMI, fever, creatinine, and number of chemo cycles (p-value > 0.05). However, in case of hematological parameters including platelets, hemoglobin, and RBCs, a significant decline was observed for the deceased group, whereas a significant heightened median expression was seen for blood urea nitrogen (BUN) (p-values < 0.01). There was a significant association observed for tumor type by survival status, wherein hematological malignancies accounted for 64% of the deaths (Table 1). Of 137 patients, 58 (42.34%) had hematologic malignancies (HM); 30 had non-Hodgkin’s lymphoma (51.7%), while 16 had acute lymphoblastic leukemia (27.5%) (Supplementary Table S2). 79 patients (57.66%) suffered from solid tumors, of these 51.8% were Breast carcinoma occurrences (supplementary Table S3).

Table 1

**Febrile Neutropenic Patient characteristics at time of hospital Admission.** Mean values (± SD) and frequency (%) for different parameters compared by survival status Alive (n = 112) Deceased (n = 25). Continuous variables for non-normal data were expressed as medians (interquartile ranges). While categorical variables were expressed as number and percentage. Comparisons between Alive and deceased groups were performed Mann Whitney rank sum test whereas Chi-square or Fisher’s exact test was employed for categorical variables.
		Alive (n = 112)			Deceased (n = 25)
Parameters		n	Column (n%)	Median (IQR)	n	Column (n%)	Median (IQR)	p-value
Gender	Female	57	50.9%		9	36.0%		0.19
Gender	Male	55	49.1%		16	64.0%		0.19
Age				36.75 (30.88, 42.45)			33.57 (31.23,43.19)	0.713
BMI				23.27 (22, 24.64)			21.83 (18.13, 26)	0.20
Fever				39 (39, 39.4)			39 (39, 39.5)	0.63
Platelets (x103 µL)				97.50 (76, 123)			13 (7.0, 54.0)	0.0001**
Haemoglobin (g/dL)				8.70 (8, 9.4)			7.7 (7.00, 8.80)	0.006**
Creatinine (mg/dl)				0.60 (0.55, 0.64)			0.61 (0.50, 0.93)	0.866
RBCs (x106 µl)				3.18 (2.95, 3.49)			2.72 (2.44, 3.22)	0.016*
Blood Urea Nitrogen (BUN) (mg/dl)				9.37 (8.05, 10.49)			13.86 (11.17,19.70)	0.0035**
No. of Chemo Cycle				2 (2, 3)			2 (2, 4)	0.76
Tumor Type	ST	70	62.5%		9	36.0%		0.023*
Tumor Type	HM	42	37.5%		16	64.0%		0.023*
Infection Type	MDI	43	38.4%		16	64.0%		.003**
	CDI	12	10.7%		5	20.0%
	FUO	57	50.9%		4	16.0%
Shock	No	108	96.4%		15	60.0%		< 0.0001***
Shock	Yes	4	3.6%		10	40.0%		< 0.0001***
Pneumonia	No	108	96.4%		16	64.0%		< 0.0001***
Pneumonia	Yes	4	3.6%		9	36.0%		< 0.0001***
ANC < 0.1	No	39	34.8%		4	16.0%		0.09
ANC < 0.1	Yes	73	65.2%		21	84.0%		0.09
Hospital Stay > 5 days	No	82	73.2%		12	48.0%		0.018*
Hospital Stay > 5 days	Yes	30	26.8%		13	52.0%		0.018*
Gram (-) Infection	No	92	82.1%		16	64%		0.05
Gram (-) Infection	Yes	20	17.9%		9	36%		0.05
IQR: Inter Quartile Range, ST: Solid Tumor, HM: Hematologic Malignancy, MDI: Microbiologically Documented Infection, CDI: Clinically Documented Infection, and FUO: Fever of Unknown Origin ANC: Absolute Neutrophil Count. p-value: * < 0.05, < 0.001, * < 0.0001

Of the neutropenic cohort, 59 (43.1%) had microbiologically documented infections (MDI), 17 (12.4%) clinically documented infections (CDI), and 61 (44.5%) fever of unknown origin (FUO) this infection type was significantly (p-value = 0.023) different between two groups. The highest occurrence of infection type amongst alive group was due to FUO (50.9%) in contrast to deceased group where major contribution towards mortality was in effect of MDI (64%) (Table 1). Out of all 24 species of bacteria identified, 16 were gram negative, while 8 were gram positive. Moreover, 2 fungal and 3 parasitic species were also identified and isolated in culture test (Supplementary table S4). 36% of deaths were attributed to gram negative infection but could not reach statistical significance (p-value = 0.058 - two tails, and 0.046 – one tail). Presence of Shock (yes = 1) (40% vs 3.6%) and pneumonia (yes = 1) (36% vs 3.6%) had a higher incidence amongst the deceased group in comparison to alive, which is indicative of a significantly strong association (p-value < 0.0001) for mortality indexing. Hospital stay > 5 days also exhibited a significant association (p-value = 0.018) with survival status (Table 1).

2. Predictive Modelling

Next, we sub-categorized clinical features (Supplementary Table S1) into three major categories (Supplementary Figure S2) towards feature selection and multivariate modelling (Supplementary Figure S3).

A. Multivariate analysis of factors underpinning FN related in-hospital mortality

To investigate patient related factors in association with in-hospital mortality, we performed categorization to enable univariate analyses (Table 1) (Supplementary Figure S2). This helped in shortlisting the first set (Set 1) of candidate predictors. We further applied multivariate analyses on these predictors of mortality and found all models to be statistically significant (p-value < 0.001, Table 2A).

B. Evaluation of chemotherapeutics in association with mortality amongst cancer patients with febrile neutropenia (FN)

Clinical guidelines indicate that certain chemotherapeutics increase the likelihood of febrile neutropenia ²⁹. To evaluate this, we isolated chemotherapeutic regimens that may confer a higher risk of FN and profiled it over survival status (Supplementary Table S5A). After primary evaluation of the chemotherapeutic regimens, the second set (Set 2) (Table 2B) of candidate predictors was collated. To evaluate the strength of these chemotherapy-related predictors towards mortality, we built three classification models (Table 2B).

C. Impact of antimicrobials and GCSF-Prophylaxis on hospital mortality

Cancer patients are provided with different types of prophylaxis as a part of their treatment including granulocyte colony-stimulating factor (GCSF)-prophylaxis and antimicrobials; antiamoebic, antiviral and antifungal prophylaxis (Supplementary Table S5B). These formed the third set of candidate predictors (Table 2C) and to determine the influence of these prophylactic treatments on mortality, we again performed predictive modelling using RF, LDA and LR (Table 2C).

3. Integrative approach towards identification of the independent predictors of hospital mortality in FN patients

To develop an integrative predictive model encompassing attributes with an enhanced predictive quality, we integrated those candidate predictors which were significant, had a high mean decrease accuracy (MDA), and had improved odds from the outcomes of individual models (Tables 1 & 2, Supplementary Figure S4 – S8, supplementary Table S6- S11). To test the effect of these selected predictors on the likelihood of patient mortality, we applied LR and found overall model to be statistically significant at, X²(12) = 57.036, p < 0.0001. The model explained 55.5% (Nagelkerke R2) of the variance in mortality and correctly classified 90.5% of cases (supplementary Table S12). Post-hoc tuning of model parameters and removal of non-contributing predictors, an adjusted model was developed. Upon bootstrapping with the over 1000 iterations, the components and characteristics of the adjusted model are summarized in (supplementary Table S13). The adjusted model was statistically significant at, X²(12) = 57.036, p < 0.0001. This integrative model explained 52.2% (Nagelkerke R²) of the variance in mortality and increased accuracy of correctly classified cases to 92%. Odds ratios for the integrative logistic model are provided in Fig. 1. To further validate the contribution of these predictors in classification of alive and dead cases, we applied RF and LDA. Outcomes from the random forest and LDA classification were in line with the logistic regression performance indices (Table 3). The five most important predictors of mortality were shock (mean decrease accuracy (MDA) = 34.07), pneumonia (27.98), hemoglobin (14.06), antiviral prophylaxis (9.82) and carboplatin (8.18) (Fig. 2). Likewise, for the LR model (Fig. 1, Supplementary S12) only shock, pneumonia and carboplatin had a significant OR. LDA provided, Fischer’s discriminant functions (DF0 (Alive) and DF1(Dead)) based on the estimation of corresponding β-values (Supplementary Table S14) which contributed towards the total variance of 100% and canonical correlation of 0.67 (p-value < 0.0001). Internal 10-fold cross validation for the integrative logistic model was performed and as a result cross validation accuracy (CVA) of 0.8608 with a kappa value of 0.437 was achieved.

Table 2

**Comparative performance indices of LR, LDA and RF models when using 3 different candidate predictor sets to predict NS associated in-hospital mortality.** The confusion matrix shows the classification of the cases based on their predicted survival status in the neutropenic dataset. Where performance indices are calculated according to : Accuracy = (TP + TN)/(TP + FP + TN + FN); sensitivity = TP/(TP + FN); specificity = TN/(TN + FP).
	(A) Predictor Set 1							(B) Predictor Set 2						(C) Predictor Set 3
	Confusion Matrix			Accuracy	Sensitivity	Specificity	p-value	Confusion Matrix		Accuracy	Sensitivity	Specificity	p-value	Confusion Matrix		Accuracy	Sensitivity	Specificity	p-value
Models	Confusion Matrix			Accuracy	Sensitivity	Specificity	p-value	Confusion Matrix		Accuracy	Sensitivity	Specificity	p-value	Confusion Matrix		Accuracy	Sensitivity	Specificity	p-value
Logistic Regression (LR)		A	D	0.883	0.938	0.64	< 0.0001	A	D	0.862	0.981	0.304	0.006	A	D	0.825	1	0.04	0.308
	A	105	7					105	2					112	0
	D	9	16					16	7					24	1
Linear Discriminant Analysis (LDA)	A	105	7	0.898	0.938	0.72	< 0.0001	105	2	0.862	0.981	0.304	< 0.0001	72	40	0.645	0.643	0.64	0.325
Linear Discriminant Analysis (LDA)	D	7	18	0.898	0.938	0.72	< 0.0001	16	7	0.862	0.981	0.304	< 0.0001	9	16	0.645	0.643	0.64	0.325
Random Forest (RF)	A	108	4	0.927	0.964	0.76	0.0002	105	2	0.862	0.981	0.261	0.14	112	0	0.818	1	0	0.55
Random Forest (RF)	D	6	19	0.927	0.964	0.76	0.0002	17	6	0.862	0.981	0.261	0.14	25	0	0.818	1	0	0.55
Predictor variables	Age, BMI, Platelets, Haemoglobin, Shock, Pneumonia, Gram Negative Infection							Cisplatin, Cyclophosphamide, Doxorubicin, Etoposide, CP, Carboplatin, Cycle						GCSF Prophylaxis, Antiamoebic Prophylaxis, Antiviral Prophylaxis, Antifungal Prophylaxis, Antibiotics 2WP, Antiamoebic 2WP, Antiviral 2WP, Antifungal 2WP
A: Alive, D: Dead, CP: Carboplatin/Paclitaxel, GCSF: Granulocyte-Colony Stimulating Factor, 2WP: Prophylaxis 2 Weeks prior hospital admission

Table 3

**Comparative performance indices of LR, LDA and RF models for final predictor set to predict NS associated in-hospital mortality.** The confusion matrix shows the classification of the cases based on their predicted survival status in the neutropenic dataset. Here, the columns denote the actual cases, and the rows denote the predicted. Where performance indices are calculated according to : Accuracy = (TP + TN)/(TP + FP + TN + FN); sensitivity = TP/(TP + FN); specificity = TN/(TN + FP).
	Final Predictor Set
	Confusion Matrix			Accuracy	Sensitivity	Specificity	p-value
Models	Confusion Matrix			Accuracy	Sensitivity	Specificity	p-value
Logistic Regression (LR)		A	D	0.92	0.955	0.76	< 0.0001
	A	105	7
	D	9	16
Linear Discriminant Analysis (LDA)	A	102	10	0.891	0.911	0.8	< 0.0001
Linear Discriminant Analysis (LDA)	D	5	25	0.891	0.911	0.8	< 0.0001
Random Forest (RF)	A	110	2	0.941	0.982	0.76	< 0.0001
Random Forest (RF)	D	6	19	0.941	0.982	0.76	< 0.0001
Predictor variables	Age, Haemoglobin, Shock, Pneumonia, Carboplatin, Doxorubicin, Antifungal 2WP, Antiviral Prophylaxis.
A: Alive, D: Dead, 2WP: 2 weeks prior prophylaxis

Nomogram for predicting in-hospital mortality

Towards clinical translation of the proposed model, a normogram was developed for prediction of mortality amongst cancer patients with FN (Fig. 3). For that, the regression coefficients for each predictor were rescaled between 0 to 100 which were then transformed into probabilities through logit transformation. Taken together, clinicians can input a patient’s profile, comprising of 8 predictors, and map them on to the nomogram to calculate the probability of a mortality. Consequently, receiver operating characteristic curve analysis gave an area under the curve of 0.878 (95% CI 0.778–0.977) (Fig. 4A). The calibration of the nomogram was checked by the calibration curve (Fig. 4B) using the Hosmer–Lemeshow test (p-value = 0.635) indicating no deviation from the fit and showing agreement between the observation and prediction. DCA conducted for the nomogram (Fig. 4C & D, Supplementary Figure S9) indicates that the proposed model may inform clinical decisions with a threshold risk probability ≥ 10%.

In this work, we present a clinical model for prognosticating outcomes for patients with FN. The proposed model was developed based on a cohort of 137 cancer patients who presented febrile neutropenia at time of hospital admission. The model predictors included bacteremia (shock, pneumonia), chemotherapeutics (carboplatin, doxorubicin), prophylaxis (antifungal and antiviral), along with hematological parameters such as hemoglobin. The resultant integrative model has a classification accuracy of 92% (Table 3) and can identify patients with febrile neutropenia (FN) at a high risk of mortality along with providing insights for prompt interventions to improve patient care and outcomes.

Cancer patients with febrile neutropenia are at a high risk of developing infections, which can rapidly overwhelm the patient and cause septic shock and death ^30–32. The current model reports shock and pneumonia to be the most important mortality predictors, with highest variable importance (Fig. 2), and odds (Fig. 1). Previously, several clinical studies have reported bacteremia and pneumonia as major causes of morbidity and mortality among patients with FN ^33,34. Bacteremic pneumonia in patients with cancer having neutropenia, is associated with a poor outcome ³³. Pseudomonas aeruginosa infection is commonly believed to primarily impact cancer patients experiencing prolonged and severe neutropenia - in particular individuals with hematologic malignancies undergoing intensive chemotherapy within hospital setting ³⁵. Moreover, in a large retrospective study, it was found that Pseudomonas aeruginosa (10%), Escherichia coli (7.7%), and Klebsiella pneumoniae (5.6%) are the leading cause for gram-negative pneumonia ³⁴. This is indicative of pneumonia being an independent predictive factor of mortality in patients ^36,37 with neutropenia.

Our analysis revealed that 64% of deaths were attributed to microbiologically documented infections (Table 1). Through various culture tests, we identified 16 types of gram-negative and 8 types of gram-positive bacterial species (Table S3). The proportion of culture positive tests were low because the patients were already on antimicrobial prophylaxis. Notably, Escherichia coli (6), Pseudomonas aeruginosa (2), and Klebsiella pneumoniae (1) infections, documented in blood culture, were associated with 9 cases of mortality. This underscores the importance of closely monitoring the clinical characteristics and risk factors associated with bacteremic pneumonia caused by these bacteria.

Moreover, thrombocytopenia is highly prevalent in ICU admittees with severe sepsis and septic shock. Its onset, whether a relative or absolute decrease in platelet count, significantly and independently correlates with a doubling of the expected mortality rate during the septic episode ^38–40. Our results show platelet counts are significantly lower in deceased patients (Table 1) and that thrombocytopenia in the ICU acts as a risk indicator, rather than a primary cause of mortality. A prompt investigation and treatment of underlying factors contributing to this condition is, therefore, clinically employable. Chemotherapy-induced thrombocytopenia (CIT) is a common complication of cancer treatment with cytotoxic agents, with carboplatin being among the most commonly implicated agents in causing CIT ^41,42 used as monotherapy or in combination with other chemo drugs. Here, we show that patients on carboplatin therapy and a neutropenic episode exhibit significant decrease in platelets (median 85 (3-125)) vs 85 (70–115) not taking carboplatin, moreover, same trend is observed for etoposide (Supplementary Table S5C). This could be indicative of drug-induced thrombocytopenia as an underlying cause that is contributing significantly towards mortality (Fig. 2) with odds increasing up to 3 times in carboplatin-treated patients.

The elevated risks associated with thrombocytopenia linked to carboplatin, etoposide, and other medications especially penicillin ⁴³, can potentiate immune-mediated thrombocytopenia and warrant further investigation. The underlying mechanism can be valuable for clinicians as it holds critical implications. In particular, for immune-mediated thrombocytopenia, avoiding the drug is imperative, whereas in dose-dependent thrombocytopenia, dose reduction may be adequate.

FN is a medical emergency, carrying a high mortality risk in the absence of timely and appropriate treatment. The use of doxorubicin and antiviral and antifungal prophylaxis plays a crucial role in reducing morbidity and mortality in patients with neutropenic sepsis ⁴⁴. Fungal pathogens are prevalent in high-risk patients experiencing neutropenia. Among these, Candida spp. and Aspergillus spp. are the most frequently implicated in invasive fungal infections ⁴⁵. Like these findings, our culture investigations also isolated these two species; therefore, provision of anti-fungal prophylaxis 2 weeks prior to the neutropenic episode provided patients with a protective effect. This is well in concordance with our model which also shows that antivirals and antifungals decrease the odds of mortality (Fig. 1 & Supplementary Table S12) and may warrant further investigation. Furthermore, the selection of antimicrobial agents for prophylaxis and empirical therapy should be driven by the local susceptibility and resistance patterns of microorganisms.

Taken together, the current model provides a novel predictor for mortality in cancer patients with FN in light of prophylactic interventions. The primary strength of our observational report is the integration of multifactorial data from a single cancer center, ensuring consistent treatment policies and standardized data collection. This integration leverages readily available prognostic factors, many of which are previously reported with insufficient statistical adjustments. Further, we applied a rigorous method in shortlisting of the potential prognostic factors from a broader range of modalities. The main limitations of this study are the retrospective nature and small size of the data. Our future work will aim to refine this model further through a validation in a larger multicenter population for more generalizability and identification of more prognostic factors. With the secondary aim of advising the integration of a more refined model into the health care system for prompt risk stratification for mortality and better implementation of the clinical management. When evaluating which method should be implemented to aid clinical review, we should mainly consider the model’s performance index, ease of implementation and interpretability. The random forests achieved the highest accuracy (94%) in identifying the occurrence of an event, however when the level of interpretation is required RF becomes difficult to interpret for an end user and is potentially more challenging to implement in real-time clinical settings. Therefore, the nomogram model was selected with an accuracy of 92% focusing on key aspects relevant to clinical usability, which was further supplemented through DCA. The model can be easily integrated into an electronic health record (EHR) system and made available as an online tool or mobile application. The model facilitates interpretable decision-making based on risk estimation and patient consultations without significant workflow disruptions.

Identification of mortality-associated factors are imperative for prognostication of cancer patients presenting neutropenia. Towards that, the factors identified in this study may offer valuable insights in resolving the underlying cause of mortality in association with shock, pneumonia, and thrombocytopenia. In particular, shock and pneumonia are reported as critical predictors, and are associated with bacteremia linked to Pseudomonas aeruginosa, Escherichia coli, and Klebsiella pneumoniae. Moreover, elevated risk is reported in association with thrombocytopenia, induced by carboplatin, etoposide, and penicillin. Combination of anti-viral and antifungal prophylaxis may improve the survival in this population. Together, a comprehensive evaluation of patient’s treatment regimen and appropriate laboratory assessments, interpreted in the clinical context, are essential for neutropenia management and in decreasing in-hospital mortality rates. Future work will investigate the use of additional treatment modalities to improve the prognostic model and risk stratification.

Conflict of interest

None, the author(s) declare no material or financial competing interests to disclose.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript and was conducted independently.

Data availability statement

The dataset generated and analyzed during the current study is not publicly available due to ongoing data analysis on the same dataset as an extension of another research project but is available from the corresponding author on reasonable request.

Moon, H., Choi, Y. J. & Sim, S. H. Validation of the Clinical Index of Stable Febrile Neutropenia (CISNE) model in febrile neutropenia patients visiting the emergency department. Can it guide emergency physicians to a reasonable decision on outpatient vs. inpatient treatment? PLoS One 13, e0210019 (2018).
Carmona-Bayonas, A. et al. Prediction of serious complications in patients with seemingly stable febrile neutropenia: Validation of the clinical index of stable febrile neutropenia in a prospective cohort of patients from the FINITE study. Journal of Clinical Oncology 33, 465–471 (2015).
de Naurois, J. et al. Management of febrile neutropenia: ESMO Clinical Practice Guidelines. Annals of Oncology 21, v252–v256 (2010).
Elting, L. S. et al. Outcomes and cost of outpatient or inpatient management of 712 patients with febrile neutropenia. J Clin Oncol 26, 606–611 (2008).
Gaieski, D. F. et al. Impact of time to antibiotics on survival in patients with severe sepsis or septic shock in whom early goal-directed therapy was initiated in the emergency department. Crit Care Med 38, 1045–1053 (2010).
Dellinger, R. P. et al. Surviving sepsis campaign: International guidelines for management of severe sepsis and septic shock: 2008. Crit Care Med 36, 296–327 (2008).
Smith, T. J. et al. 2006 update of recommendations for the use of white blood cell growth factors: an evidence-based clinical practice guideline. J Clin Oncol 24, 3187–3205 (2006).
Herbst, C. et al. Prophylactic antibiotics or G-CSF for the prevention of infections and improvement of survival in cancer patients undergoing chemotherapy. Cochrane Database of Systematic Reviews (2009) doi:10.1002/14651858.CD007107.PUB2/MEDIA/CDSR/CD007107/REL0002/CD007107/IMAGE_T/TCD007107-AFIG-FIG01.PNG.
Shmuely, H., Monely, L. & Shvidel, L. All-Cause Mortality and Its Predictors in Haemato-Oncology Patients with Febrile Neutropenia. J Clin Med 12, (2023).
Carmona-Bayonas, A. et al. Prediction of serious complications in patients with seemingly stable febrile neutropenia: Validation of the clinical index of stable febrile neutropenia in a prospective cohort of patients from the FINITE study. Journal of Clinical Oncology 33, 465–471 (2015).
Klastersky, J. et al. The Multinational Association for Supportive Care in Cancer Risk Index: A Multinational Scoring System for Identifying Low-Risk Febrile Neutropenic Cancer Patients. https://doi.org/10.1200/JCO.2000.18.16.3038 18, 3038–3051 (2016).
Kuderer, N. M., Dale, D. C., Crawford, J., Cosler, L. E. & Lyman, G. H. Mortality, morbidity, and cost associated with febrile neutropenia in adult cancer patients. Cancer 106, 2258–2266 (2006).
Nordvig, J. et al. Febrile Neutropenia and Long-term Risk of Infection Among Patients Treated With Chemotherapy for Malignant Diseases. Open Forum Infect Dis 5, (2018).
Lalami, Y. & Klastersky, J. Impact of chemotherapy-induced neutropenia (CIN) and febrile neutropenia (FN) on cancer treatment outcomes: An overview about well-established and recently emerging clinical data. Crit Rev Oncol Hematol 120, 163–179 (2017).
Du, X. et al. Predicting in-hospital mortality of patients with febrile neutropenia using machine learning models. Int J Med Inform 139, (2020).
Green, M. D. et al. A randomized double-blind multicenter phase III study of fixed-dose single-administration pegfilgrastim versus daily filgrastim in patients receiving myelosuppressive chemotherapy. Annals of Oncology 14, 29–35 (2003).
Lyman, G. H. & Delgado, D. J. Risk and timing of hospitalization for febrile neutropenia in patients receiving CHOP, CHOP-R, or CNOP chemotherapy for intermediate-grade non-Hodgkin lymphoma. Cancer 98, 2402–2409 (2003).
Crawford, J. et al. Risk and timing of neutropenic events in adult cancer patients receiving chemotherapy: the results of a prospective nationwide study of oncology practice. J Natl Compr Canc Netw 6, 109–118 (2008).
Lyman, G. H., Abella, E. & Pettengell, R. Risk factors for febrile neutropenia among patients with cancer receiving chemotherapy: A systematic review. Crit Rev Oncol Hematol 90, 190–199 (2014).
Aagaard, T. et al. Mortality and admission to intensive care units after febrile neutropenia in patients with cancer. Cancer Med 9, 3033 (2020).
Requirements (2018 Common Rule) | HHS.gov. https://www.hhs.gov/ohrp/regulations-and-policy/regulations/45-cfr-46/revised-common-rule-regulatory-text/index.html#46.116.
Mehta, H. M., Malandra, M. & Corey, S. J. G-CSF and GM-CSF in Neutropenia. The Journal of Immunology 195, 1341–1349 (2015).
Giri, R. K. & Sahoo, R. K. Febrile Neutropenia. Onco-critical Care: An Evidence-based Approach 233–250 (2023) doi:10.1007/978-981-16-9929-0_21.
Erdfelder, E., FAul, F., Buchner, A. & Lang, A. G. Statistical power analyses using G*Power 3.1: Tests for correlation and regression analyses. Behav Res Methods 41, 1149–1160 (2009).
Shipe, M. E., Deppen, S. A., Farjah, F. & Grogan, E. L. Developing prediction models for clinical use using logistic regression: an overview. J Thorac Dis 11, S574–S584 (2019).
Zhang, Z. & Kattan, M. W. Drawing Nomograms with R: applications to categorical outcome and survival data. Ann Transl Med 5, (2017).
Harrell, F. E. Regression Modeling Strategies. (2001) doi:10.1007/978-1-4757-3462-1.
Towards trustable machine learning. Nat Biomed Eng 2, 709–710 (2018).
Okunaka, M., Kano, D., Matsui, R., Kawasaki, T. & Uesawa, Y. Comprehensive Analysis of Chemotherapeutic Agents That Induce Infectious Neutropenia. Pharmaceuticals 14, (2021).
Marín, M. et al. Factors influencing mortality in neutropenic patients with haematologic malignancies or solid tumours with bloodstream infection. Clinical Microbiology and Infection 21, 583–590 (2015).
Ramzi, J. et al. Predictive factors of septic shock and mortality in neutropenic patients. Hematology 12, 543–548 (2007).
Clarke, R. T., Jenyon, T., Parsons, V. V. H. & King, A. J. Neutropenic sepsis: management and complications. Clinical Medicine 13, 185 (2013).
Carratalà, J., Rosón, B., Fernández-Sevilla, A., Alcaide, F. & Gudiol, F. Bacteremic Pneumonia in Neutropenic Patients With Cancer: Causes, Empirical Antibiotic Therapy, and Outcome. Arch Intern Med 158, 868–872 (1998).
Zhao, C. et al. Risk Factors for 30-Day Mortality in Patients with Bacteremic Pneumonia Caused by Escherichia coli and Klebsiella pneumoniae: A Retrospective Study. Int J Gen Med 16, 6163 (2023).
Maschmeyer, G. & Braveny, I. Review of the incidence and prognosis of Pseudomonas aeruginosa infections in cancer patients in the 1990s. Eur J Clin Microbiol Infect Dis 19, 915–925 (2000).
Chang, H. Y. et al. Causes of death in adults with acute leukemia. Medicine (United States) 55, 259–268 (1976).
Evans, S. E. & Ost, D. E. Pneumonia in the neutropenic cancer patient. Curr Opin Pulm Med 21, 260 (2015).
Venkata, C., Kashyap, R., Christopher Farmer, J. & Afessa, B. Thrombocytopenia in adult patients with sepsis: Incidence, risk factors, and its association with clinical outcome. J Intensive Care 1, 1–10 (2013).
Larkin, C. M., Santos-Martinez, M. J., Ryan, T. & Radomski, M. W. Sepsis-associated thrombocytopenia. Thromb Res 141, 11–16 (2016).
Awad, W. B., Nazer, L., Elfarr, S., Abdullah, M. & Hawari, F. A 12-year study evaluating the outcomes and predictors of mortality in critically ill cancer patients admitted with septic shock. BMC Cancer 21, 1–7 (2021).
Wu, Y. H., Chen, H. Y., Hong, W. C., Wei, C. Y. & Pang, J. H. S. Carboplatin-Induced Thrombocytopenia through JAK2 Downregulation, S-Phase Cell Cycle Arrest and Apoptosis in Megakaryocytes. Int J Mol Sci 23, (2022).
Ten Berg, M. J. et al. Thrombocytopenia in adult cancer patients receiving cytotoxic chemotherapy: Results from a retrospective hospital-based cohort study. Drug Saf 34, 1151–1160 (2011).
Bakchoul, T. & Marini, I. Drug-associated thrombocytopenia. Hematology 2018, 576–583 (2018).
Villafuerte-Gutierrez, P., Villalon, L., Losa, J. E. & Henriquez-Camacho, C. Treatment of Febrile Neutropenia and Prophylaxis in Hematologic Malignancies: A Critical Review and Update. Adv Hematol 2014, (2014).
Pfaller, M. A., Pappas, P. G. & Wingard, J. R. Invasive Fungal Pathogens: Current Epidemiological Trends. Clinical Infectious Diseases 43, S3–S14 (2006).

There is NO Competing Interest.

Supplementary02.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

In-hospital Mortality Prognostication for Cancer Patients with Febrile Neutropenia: A Single Center Observational Study

Status:

Version 1

Abstract

Figures

Introduction

Patients and Methods

Setting and study population

Data collection and Prognostic factors

Statistical analysis

Logistic Regression

Interpretability

Results

1. Patient Characteristics

2. Predictive Modelling

A. Multivariate analysis of factors underpinning FN related in-hospital mortality

B. Evaluation of chemotherapeutics in association with mortality amongst cancer patients with febrile neutropenia (FN)

C. Impact of antimicrobials and GCSF-Prophylaxis on hospital mortality

3. Integrative approach towards identification of the independent predictors of hospital mortality in FN patients

Nomogram for predicting in-hospital mortality

Discussion

Conclusion

Declarations

Conflict of interest

Funding

Data availability statement

References

Additional Declarations

Supplementary Files

Status:

Version 1