Phenotyping Multiple Organ Dysfunction Syndrome Using Temporal Trends in Critically Ill Children

Abstract

Multiple organ dysfunction syndrome (MODS) is one of the most common causes of death in critically ill children. However, despite decades of clinical trials, there are no comprehensive approaches to the management of MODS or effective targeted therapies that have consistently improved outcomes. Better understanding the heterogeneity of MODS and characterizing subgroups of MODS patients could improve our understanding of the syndrome and help us develop new management strategies. We analyzed a cohort of 5,297 children with MODS from two children’s hospitals and used subgraph-augmented non-negative matrix factorization (SANMF) to identify unique temporal patterns in organ dysfunction across four novel subgroups. We demonstrate that these subgroups are composed of patients with distinct clinical characteristics and are independently predictive of clinical outcomes. Our work suggests that these subgroups represent four relevant phenotypes of pediatric MODS that could be used to identify novel management strategies.

I. Introduction

The development of multiple organ dysfunction syndrome (MODS) is a common final pathway for death in critically ill children [1]. MODS is a complex heterogeneous syndrome that can develop after many types of injury (e.g. sepsis, trauma, major surgery) and is made of subgroups of patients with distinct phenotypes (e.g. different underlying molecular pathways, types of organ dysfunctions, etc.) but these phenotypes remain ill-defined and the therapeutic options limited [1], [2]. Better characterizing the phenotypes of pediatric MODS could support the development of more effective therapies and prognostic models.

In this study, we aimed to identify and characterize novel phenotypes of pediatric MODS with prognostic implications by analyzing the dynamic trends of individual organ dysfunctions over time. We used the Pediatric Sequential Organ Failure Assessment (pSOFA) score to characterize the type and severity of organ dysfunction [3]. Organ dysfunction is sometimes not apparent until days after the onset of illness, therefore, we used the pSOFA subscores from the first three days after PICU admission, rather than a single time point, to characterize the trends in organ dysfunctions using subgraph mining [4]. We then used non-negative matrix factorization (NMF) to cluster patients into subgroups based on these temporal patterns of organ dysfunction using the extracted subgraphs, [5] and characterized the patient subgroups based on their demographics, disease severity, and outcomes. Finally, we showed that these subgroups are important predictors of clinical outcomes including mortality, persistent MODS at day 7, and length of stay.

II. Related Work

Previous work focused on clustering critically ill patients based on dynamic trends in clinical data have used two main approaches. First, temporal-based techniques have been used to analyze times series of physiological variables to identify relevant clusters of patients [6], [7]. Second, recent work has focused on detecting trends in complex time series data utilizing recurrent neural networks (RNNs) [8]. These studies, although they present promising results, sacrifice interpretability to improve accuracy and other error metrics. To address these challenges, matrix factorization has been used to group temporal trends while maintaining clinical meaningfulness. Luo et al. developed and applied specialized NMF algorithms to cluster patients with a variety of conditions [11], [12], [13]. Specifically, subgraph-augmented NMF converts physiological times series data into graphs that describe a patient’s disease progression over their hospital admission [12]. The frequency of individual subgraphs within the dataset is utilized to identify clusters of patients with common disease trajectories with maximum interpretability. Here, we apply these techniques to pSOFA scores to identify unique temporal trends underlying MODS progression in children.

III. Methods

A. Collecting Data and Generating Graphs

Data for this study were collected from the PICUs at both Ann and Robert H. Lurie Children’s Hospital of Chicago and the University of Chicago Comer Children’s Hospital for admissions between 1/2010 and 8/2016. Laboratory measurements for bilirubin (mg/dL), creatinine (mg/dL), mean arterial pressure (mmHg), platelets (count × 10³/μL), and vasoactive infusion of dopamine hydrochloride, dobutamine hydrochloride, epinephrine, or norepinephrine bitartrate (μg/kg/min) were collected for each patient. Additionally, the arterial blood gas ratio (PaO₂:FiO₂) or peripheral oxygen saturation (SpO₂:FiO₂) were recorded and the Glasgow Coma Score was measured on the pediatric scale. Each clinical measurement included the time of collection relative to PICU admission. Missing variables were considered normal and the corresponding pSOFA subscore was assigned to zero. Age-adjusted pSOFA subscores were summarized at the day-level using the maximum score for each 24-hour period..

Severity of illness on admission was determined by the Pediatric Risk of Mortality (PRISM) III score using physiological variables from the first 24 hours [14]. Chronic comorbidities were based on Feudtner’s classification [15]. Patients with immunocompromise state included those with an oncologic disease and the recipients of a solid or stem cell transplant. Patients who had antibiotics administered and microbiological cultures obtained during the first 72 hours were considered to have a confirmed or suspected infection.

MODS was defined as having a subscore of 2 or more in 2 or more organ systems based on the pSOFA score in the first 72 hours. Out of 20,827 patients admitted to both PICUs, 5,297 had MODS and were included in the analysis.

B. Subgraph Mining

Node-edge list graphs were prepared for each of the 5,297 patients with MODS. Common subgraphs were extracted from the pool of graphs generated for each patient using the Molecular Substructure Miner (MoSS) developed by Christian Borgelt [4]. The MoSS search algorithm was run with no complement group and with no frequency requirement to mine as many potential subgraphs as possible. We identified 528 distinct subgraphs with at least two nodes using just three data points per patient for each of the 6 organ systems quantified by pSOFA subscores.

The matrix of patient-subgraph counts was processed to reduce redundancy by counting only the largest subgraph present and not any smaller subgraphs contained within, as described by [12]. 106 subgraphs were removed from the dataset in this process. Additionally, subgraphs that were static and did not reach a maximum subscore of 3 or above were removed from analysis (n = 36). Subgraphs occurring in fewer than 20 patients were excluded (n = 241). Finally, patients who no longer had any quantified subgraphs after data cleaning were excluded from downstream analysis (n = 243). The final dataset contained information on 145 subgraphs for 5054 patients. Of the 145 subgraphs used to identify different patient trajectories, 31 depicted escalations in organ subscores or deterioration in patient status, 24 represented severe subscores with no change over time, and 90 subgraphs described improvement in organ subscores.

C. Sparse NMF Implementation

The cleaned dataset was then partitioned into training and testing splits at an 80%/20% ratio, stratified by mortality. We confirmed that the testing set was representative of the training set based on clinical variable distribution. The number of subgroups was determined on the training set using the cophenetic correlation, Akaike information criterion (AIC), and clinical judgment on the training set. Non-negative matrix factorization was implemented enforcing sparsity using the SNMF/L algorithm with Nonnegative Double Singular Value Decomposition (nndsvd) seeding as described in [5], [11], [12].

D. Identifying and Characterizing MODS Subgroups

Representative subgraphs were chosen such that selected subgraphs accounted for the top 75% of group coefficients. We limited the number of subgraphs per group to 10 to maintain interpretability. We examined the frequency of each subgraph within its group to determine the relative importance to the identified disease pattern. Patients were assigned group membership based on the highest probability group in the mixture matrix from the initial NMF model. Additionally, we compared patient outcome variables across groups to better understand the clinical manifestation of different MODS phenotypes.

To internally confirm the validity of our selected subgraphs, we carried out a multinomial logistic regression predicting group membership using the minimal set of subgraphs in the testing set. After confirming the subset of subgraphs was sufficient for explaining the different groups, we performed functional validation. We examined the importance of group membership in predicting patient outcomes via linear and logistic regression after adjusting for severity of illness, age, immunocompromised status, and study site. Finally, we considered the clinical relevance of these groups by contrasting common laboratory and outcome measurements and ensuring that trends identified in the training set were recapitulated in the testing set using adjusted logistic regression.

IV. Results

A. Pediatric MODS patients can be separated into 4 distinct subgroups based on organ dysfunction trajectory within the first 72 hours.

Utilizing sparse non-negative matrix factorization (SNMF), we identified four distinct groups that describe changes in patient pSOFA subscores over the first 72 hours in the PICU. For each group, we selected the subgraphs that described the top 75% of group coefficients, with a maximum of 10 subgraphs used per group. Of the 40 subgraphs chosen, 32 exclusively describe one group, indicating the uniqueness of the groups identified.

To assess the validity of describing each group with only 8–10 subgraphs, we performed a multinomial logistic regression predicting subgroup membership using only the 33 representative subgraphs and compared the results to the SNMF subgroup assignments. This simplified model agreed with SNMF classifications 89.0% of the time in the training group and 88.1% in the testing group. Additionally, the group-wise precision and recall scores are shown in Table 1. These results indicate that the small number of selected subgraphs stratify patients into the four subgroups well when compared to the SNMF clustering using all subgraphs.

TABLE I.

Error metrics for multinomial-logistic regression

Measure	Split	Subgroup 1	Subgroup 2	Subgroup 3	Subgroup 4
Precision	Train	0.804	0.854	0.910	1.000
	Test	0.788	0.844	0.915	0.996
Recall	Train	1.000	0.998	1.000	0.684
	Test	0.983	0.994	0.989	0.678
F-score	Train	0.891	0.920	0.953	0.812
	Test	0.875	0.913	0.951	0.807

B. Patients in MODS subgroups have variable demographic and disease characteristics.

To determine whether the four groups identified through SNMF represent differences in patient trajectories and pSOFA scores, we plotted the average pSOFA score for each group over the first 7 days after PICU admission (Figure 1B). Subscores are sorted by magnitude in these plots, showing that Subgroup 1 is driven by a high neurologic subscore (signifying severe encephalopathy), Subgroup 3 is driven by high respiratory subscores (or severe hypoxemia), and Subgroup 4 is driven by high cardiovascular and coagulation subscores (or shock and thrmobocytopenia). In comparison, Subgroups 2 exhibits lower pSOFA scores.

Fig. 1.

Pediatric MODS subgroups are defined by unique temporal patterns in pSOFA subscores. (A) Representative subgraphs that define the four subgroups, labeled with the subgraph ID number and frequency within the subgroup in parentheses. Subgraphs were chosen according to the top 75% of group coefficients and limited to no more than 10 per subgroup. (B) Average pSOFA scores by organ system within each subgroup for 7days after PICU admission.

As shown in Table 2, clinical characteristics, laboratory values, interventions, and outcomes were comparable within subgroups between the training and testing sets and distinct across subgroups.

TABLE II.

Comparison of laboratory measurements and clinical characteristics across subgroups and data splits

MEASUREMENTa	SUBGROUP 1		SUBGROUP 2		SUBGROUP 3		SUBGROUP 4
	Training (n = 802)	Testing (n = 217)	Training (n = 1463)	Testing (n = 365)	Training (n = 823)	Testing (n = 189)	Training (n = 956)	Testing (n = 239)
Age (months)	48.9 [11.7, 142.7]	58.3 [12.6, 147.5]	48.6 [11.0, 138.1]	42.6 [11.3, 126.6]	41.0 [9.8, 135.1]	45.6 [9.6, 139.0]	101.8 [22.6, 182.9]	78.1 [19.1, 178.0]
Min. Hemoglobin (g/dL)	9.2 [7.5, 11.0]	8.9 [7.2, 10.8]	9.5 [7.8, 11.2]	9.9 [7.7, 11.6]	8.8 [7.2, 10.5]	8.8 [7.2, 10.5]	8.3 [6.9, 10.0]	8.3 [7.1, 10.2]
Min. Platelet Count (count x 103/μL)	186.0 [107.0, 267.0]	178.0 [101.0, 271.8]	188.0 [95.0, 272.0]	188.0 [90.8, 267.3]	155.0 [68.0, 243.5]	158.0 [68.0, 269.0]	84.0 [28.0, 180.5]	78.0 [23.5, 176.0]
Min. Lymphocyte Count (count x 103/μL)	1.3 [0.7, 2.2]	1.2 [0.8, 2.2]	1.3 [0.7, 2.2]	1.4 [0.7, 2.4]	1.1 [0.5, 1.9]	1.2 [0.5, 2.1]	0.8 [0.3, 1.5]	0.8 [0.2, 1.5]
Max. Bilirubin (g/dL)	0.5 [0.3, 1.2]	0.4 [0.2, 1.0]	0.6 [0.2, 2.3]	0.4 [0.3, 1.4]	0.6 [0.2, 1.5]	0.6 [0.3, 1.7]	1.2 [0.5, 3.0]	1.2 [0.5, 4.1]
Min. Albumin (g/dL)	3.1 [2.5, 3.6]	3.1 [2.5, 3.6]	3.1 [2.6, 3.7]	3.2 [2.7, 3.8]	2.7 [2.2, 3.3]	2.6 [2.2, 3.2]	2.8 [2.3, 3.4]	2.8 [2.2, 3.4]
Length of stay after PICU admission	10.0 [4.7, 20.9]	9.0 [4.3, 19.5]	8.2 [4.6, 15.6]	7.8 [4.1, 14.2]	11.3 [4.9, 22.4]	11.1 [4.7, 21.9]	10.0 [5.8, 18.6]	9.9 [5.5, 19.9]
Suspected or Confirmed Infection by day 3	368 (45.9)	104 (47.9)	642 (43.9)	159 (43.6)	527 (64.0)	123 (65.1)	565 (59.1)	143 (59.8)
Immunocompromised	79 (9.9)	26 (12.0)	194 (13.3)	46 (12.6)	130 (15.8)	24 (12.7)	300 (31.4)	82 (34.3)
Vasoactive Med. by day 3	181 (22.6)	46 (21.2)	185 (12.6)	46 (12.6)	360 (43.7)	86 (45.5)	504 (52.7)	122 (51.0)
Mech. Vent. by day 3	730 (91.0)	196 (90.3)	1136 (77.6)	273 (74.8)	715 (86.9)	165 (87.3)	429 (44.9)	124 (51.9)
PRISM III (Pediatric Risk of Mortality score)	8.0 [5.0, 15.0]	9.0 [4.0, 14.0]	6.0 [2.0, 10.0]	5.0 [2.0, 10.0]	10.0 [5.0, 19.0]	10.0 [5.0, 19.0]	8.0 [5.0, 13.0]	9.0 [5.0, 14.5]
Persistent MODS by day 7	269 (33.5)	85 (39.2)	233 (15.9)	58 (15.9)	336 (40.8)	81 (42.9)	255 (26.7)	82 (34.3)
Died in hospital	88 (11.0)	28 (12.9)	47 (3.2)	13 (3.6)	145 (17.6)	31 (16.4)	69 (7.2)	25 (10.5)

^aFor all measurements, max. = maximum and min. = minimum. Continuous variables are shown as ”median [IQR]”. Categorical variables are ”Positive count (percent)”.

^bWhite blood cell count;

^cAlanine aminotransferase

Based on the disease trajectories indicated by representative subgraphs, descriptors of disease severity, and patient laboratory measurements, we described each group as follows:

• Subgroup 1 - Severe encephalopathy with moderate organ dysfunction

• Subgroup 2 - Moderate resolving hypoxemia

• Subgroup 3 - Severe persistent hypoxemia and shock

• Subgroup 4 - Persistent cytopenias, hepatobiliary dysfunction, and shock

C. MODS subgroups are independently associated with clinical outcomes.

In addition to characterizing the differences between groups, we wanted to determine if subgroup membership was associated with in-hospital mortality, persistent MODS at day 7, and length of stay after adjusting for other confounders of outcomes, including age, severity of illness by PRISM III score, immunocompromised status, and study site. We performed multivariate logistic regression and found that subgroup membership was independently associated with mortality (Type II Anova, p-value = 5.831e-07) and MODS at day 7 (Type II Anova, p-value = 1.116e-14). Additionally, we carried out a linear regression for length of stay and again found subgroup was independently associated (Type II Anova, p-value = 3.533e-3). These findings indicate the subgroups identified through SNMF not only separate patients based on disease characteristics within the first 72 hours, but are also associated with outcomes independent of other usual confounders.

V. Discussion

We have shown that pediatric MODS patients can be split into four distinct subgroups based on dynamic trends of organ dysfunction. These subgroups exhibit different clinical characteristics, associated laboratory values, disease severity, and are reproducible in a testing set. Furthermore, these subgroups are independently associated with important clinical outcomes, such as persistent MODS at 7 days, length of stay, and in-hospital mortality. We show here that these subgroups are prognostically relevant, but it is possible that this subgrouping schema could have therapeutic implications. It is likely that many of the patients within each subgroup share many of the same underlying pathobiological disturbances, as evidenced by the shared clinical characteristics and laboratory values. Understanding the common features of these underlying pathobiologies could help determine which targeted therapies could be most beneficial to individual patients [16]

Several researchers have previously used data-driven approaches to identify subgroups of adult patients with sepsis and sepsis-induced MODS using clinical data [17], [18]. Knox et al. [17] employed self-organizing maps in a cohort of 2,533 critically ill adult patients with severe sepsis or septic shock and identified four clusters. Upon comparison, our ”severe persistent hypoxemia and shock” subgroup had many similarities to their ”shock with hypoxemia and altered mental status” cluster, and our ”persistent cytopenias, hepatobiliary dysfunction, and shock” subgroup had similarities to their ”hepatic disease” cluster, in terms of coagulopathy and liver involvement. In both cases, these were two of the phenotypes associated with higher mortality rates [17]. More recently, Seymour et al. [18] identified four phenotypes of sepsis based consensus clustering of clinical data in a cohort of 20,189 adult patients. Their γ phenotype shared many similarities with our ”severe persistent hypoxemia and shock” subgroup, in terms of hypoxemia, vasopressor need, low albumin, and high mortality. In addition, their δ phenotype was similar to our ”persistent cytopenias, hepatobiliary dysfunction, and shock” subgroup based on the coagulation dysfunction, liver involvement, and higher mortality. Both of these adult studies only used data from presentation to derive the sepsis phenotypes, whereas we used temporal trends over the first 3 days to characterize our MODS subgroups. To our knowledge, our study is the first one to use a data-driven approach to derive pediatric MODS subgroups using clinical data.

Our findings are subject to several limitations that could be addressed in future studies. First, the data for this study were collected from only two sites in close geographical proximity, therefore the conclusions herein may not represent the entire continuum of pediatric organ dysfunction. Second, the subgraphs that characterize four distinct organ dysfunction trajectories were mined with single-day resolution. Utilizing a 24-hour window may mask shorter trends that, while significant for MODS development or resolution, are not present for entire days at a time. This could be addressed in future studies by using a higher temporal resolution. Likewise, we limited the time period of interest to within 72 hours of PICU admission, but this could be expanded to include data over more days of hospitalization to explore trends in longer time windows.

VI. Conclusion

We uncovered and characterized four unique, prognostically relevant, and reproducible subgroups of pediatric MODS using temporal patterns of organ dysfunction in critically ill children. These subgroups exhibit distinct clinical characteristics and outcomes independent of age, severity of illness, immunocompromised status, and study site. These findings could be leveraged to design targeted therapies and identify management strategies that are most likely to benefit specific subgroups patients. Further work is needed to validate these findings in larger, multi-center cohort of critically ill children and evaluate their clinical utility.