Introduction
Septic shock is a leading cause of mortality and morbidity in the intensive care unit (ICU). The shock status should be corrected as soon as possible to prevent subsequent development of multiple organ dysfunctions [
1‐
3]. The resuscitation of septic shock at the initial phase involves fluid infusion and use of vasoactive agents such as norepinephrine, dopamine and dobutamine [
4,
5]. Although the Surviving Sepsis Campaign guidelines recommend several goals (i.e., urine output, mean blood pressure and ScvO
2) to guide resuscitation [
6,
7], the specific strategy must be individualized because the responses to a given intervention can vary greatly among septic shock patients. For example, some patients with sepsis-induced acute kidney injury may respond well to fluid challenge, while others may further develop renal failure after fluid resuscitation [
8,
9]. The clinical heterogeneity must be accounted for in both clinical practice and clinical trial design. Since sepsis and/or septic shock is a heterogeneous clinical syndrome, many clinical trials targeting sepsis population usually result in neutral findings [
2,
10,
11]. In these trials, some patients may benefit from a certain intervention, but others will be harmed by the intervention, resulting in a neutral effect in the overall population. Thus, numerous efforts have been made to explore the heterogeneity of sepsis.
Sepsis has been found to be consisted of several phenotypes, though specific class membership assignments are different across studies [
12‐
18]. By using clinical trial data from 1696 patients, Gårdlund B and colleagues identified six classes of septic shock [
19]. Alternatively, the use of temperature trajectory was able to identify four classes of sepsis [
20]. The identified classes were investigated for their responses to different treatments. Our previous results showed that these clinical subclasses have different responses to the amount of fluid infusion [
14]. By utilizing randomized clinical trial data, the proportion of RCTs reporting benefit, harm or no effect changed considerably by varying the proportion of subclasses [
21]. These results indicated that the subclasses of sepsis should be considered in designing clinical trials because of their differing responses to fluid strategies. However, previous studies primarily explored subclasses of sepsis using cross-sectional data, ignoring the transition pattern of classes and the sequential treatment decisions. In real clinical practice, septic shock is managed with sequential treatment decisions. In other words, the treatment decision making in the current stage should consider not only the current status but also previous responses to the treatment. Unfortunately, such a clinical practice pattern has not been formalized with mathematical modeling.
Previous studies have explored the feasibility of utilizing high-granularity dataset to develop sequential decision rules of resuscitation for septic shock. For example, Komorowski M and colleagues developed a reinforcement learning algorithm to determine the sequential rules of treatment strategy [
22]. Our study group utilized dynamic treatment regime (DTR) model to develop a sequential treatment strategy [
23]. However, these models are of high granularity with limited explainability. To make the model more explainable, the present study firstly classified septic shock into several classes by using non-supervised learning algorithms. Then, the classification system was integrated in a DTR model to develop a sequential treatment rule for fluid volume and vasopressor dosing [
24]. The optimal treatment strategy was compared with the actual strategy through days 0–7, and risk factors for fluid overload and norepinephrine overdosing were explored. We hypothesized that several classes of septic shock could be robustly identified and DTR model was able to identify optimal treatment strategy for these classes. The differences between actual and optimal treatment strategy varied across different classes.
Discussion
The study identified five classes of septic shock that showed distinct clinical characteristics by using finite mixture modeling, which was further confirmed by k-means clustering. DTR model was employed to tailor individualized sequential treatment decisions, by estimating the optimal fluid volume and norepinephrine dosing. The five classes are clinically relevant in that (1) they are easily identifiable by routine clinical variables with stable prediction probability (minimum class membership probability > 0.80); (2) the optimal resuscitation strategy, which was confirmed in an independent dataset, differed across the five classes; (3) the transition from resuscitation to de-resuscitation phase should be different across classes so as to achieve a desirable clinical outcome; and (4) these classes showed differing risks of fluid overloading and norepinephrine overdosing. The system has important clinical implications that the heterogeneous septic shock population can be classified into subphenotypes and resuscitation strategy can be tailored by considering subphenotypes as well as their transitions across ICU days.
This study confirms previous findings that septic shock is a heterogeneous syndrome and can be classified into several stable subclasses [
14,
19,
20]. The classification is stable in our multicenter cohort because both finite mixture modeling and
k-means clustering arrive at the same number of classes. The minimum class membership probability is greater than 0.80. This study is different from previous studies that we focus solely on septic shock because we believe that this population requires urgent resuscitation and can benefit most from individualized treatment regimen. By using multicenter clinical data, Seymour CW and colleagues identified four types of sepsis, namely the
\(\alpha\),
\(\beta\),
\(\gamma\),
\(\delta\) phenotypes [
21]. Septic shock is mostly in the
\(\delta\) phenotype, which is also associated with the highest mortality rate. From the perspective of immune responses, septic shock was classified into three subclasses in another study, but the small sample size prohibited finer classification [
38]. Gårdlund and colleagues reported similar classes of septic shock. For example, the “uncomplicated septic shock” profile corresponds to class 5 (mild class) in our study, and the “severe septic shock” profile corresponds to class 2 (critical class) [
19].
In the DTR framework [
3439,
4041], the optimal sequential resuscitation rules were estimated. Days 0–3 were considered as resuscitation phase, and day 7 was considered as the de-resuscitation phase. Consistent with the concept of the “four D's” of fluid therapy [
42,
43], our DTR model showed that larger fluid infusion and appropriate dosing of norepinephrine were usually required to achieve a better clinical outcome at an early phase, and less fluid infusion was beneficial at the late phase. Our study also showed that specific dosing strategies were different among classes. For example, the de-resuscitation phase began on day 3 for class 1 but began on day 1 for class 3 (renal failure class). Class 3 is at increased risk of fluid overload because the injured kidney is unable to effectively maintain fluid balance. Thus, patients in this class are more sensitive to fluid therapy, which is supported by rapid drop in optimal fluid volume from day 0 to day 1. Furthermore, class 3 is more likely to transition to other classes as it is the largest class at day 0, but the size decreases rapidly over time (Fig.
2C). The actual fluid volume was relatively low in class 4 (respiratory failure class), suggesting that physicians are aware of potential hazardous effect of fluid overloading for injured lungs [
44‐
47].
There are numerous tools to evaluate fluid overload in clinical practice such as physical examination, chest radiography, natriuretic peptides, thoracic ultrasound and bioelectrical impedance analysis [
48]. But these methods are inaccurate and variable across individuals. This study estimated the optimal fluid volume in the framework of DTR, which integrated many relevant clinical variables for model training, allowing for individualized fluid treatment strategy. The DTR models were also well validated in an independent dataset. Some interesting risk factors such as body weight, urine output and PaCO
2 were identified for fluid overloading. The body weight may not be a good marker to determine fluid dose because fluid retention is common in critically ill patients. Increased body weight is a sign of fluid overload, but in reality, physicians may prescribe too much fluid based on the formula for calculating fluid requirement. Furthermore, patients with high PaCO
2 are more likely to have fluid overload because PaCO
2 retention can be the result of severe ARDS and protective ventilation, in which the optimal fluid is usually conservative in order to improve clinical outcomes.
Norepinephrine is usually required to maintain blood pressure after fluid infusion. However, the timing and dosing of norepinephrine are largely based on subjective judgment. As compared to the real clinical practice in our participating hospitals, less norepinephrine and more fluids can be given to class 2 patients on day 0, in order to achieve a better clinical outcome. This is consistent with some observations that treatment strategy with more fluid volume and lower vasopressor dose at 0–6 h is associated with improved mortality outcome [
49,
50]. Patients in class 2 are characterized by profound hypotension and poor tissue perfusion. The use of norepinephrine may further reduce tissue perfusion when fluid infusion is inadequate. It is interesting to note that the norepinephrine dosing is similar between actual and optimal strategy on day 1, which supports the well-accepted concept that adequate fluid infusion should be given before considering vasopressors. Results from observational studies showed that early norepinephrine use is associated with improved mortality outcome [
51‐
54], which is further supported by a small RCT [
55]. This is not contradictory to our results. The actual norepinephrine is underdosed on days 0 and 1 in classes 1 and 3 in our study. In other words, early larger dose of norepinephrine in these two classes could help improve outcomes. Since classes 1 and 3 comprise the majority of septic shock population, it is not surprising that the undifferentiated septic shock patients enrolled in those trials can benefit from early vasopressor use.
There are several limitations that must be acknowledged in the study. Although we tried to classify septic shock patients into five subclasses, the granularity may not be high enough to fully implement the individualized resuscitation strategy. Sepsis is a highly heterogeneous and dynamic syndrome, and it is likely that different patients within each of the clusters will require different resuscitation regimen. However, higher granularity means the requirement of large sample size and makes the explainability of the model more challenging. We need to strike a balance between granularity, sample size requirement and explainability. Secondly, there could be residual confounding effect in the Cox regression and the DTR models due to the observational nature of the study design. For example, the study showed that larger doses of norepinephrine were consistently associated with increased risk of mortality, which could be explained by the fact that patients requiring larger dose of norepinephrine were more critically ill. Thirdly, the minimum number of patients was predefined to ensure each cluster contained clinically meaningful size, so that optimal treatment strategy could be explored within each cluster. However, it is possible that by forcing patients into larger/fewer clusters, patients with potentially different treatment responses are being lumped together. This is actually a trade-off between high granularity and model explainability. Finally, it is ideal to explore the fluid and norepinephrine dosing in the same model. However, the unique combinations of treatments can result in large number of interventions, which is not allowed with limited sample size.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.