Low Quality Evidence Supporting Recommendations in the 2021 Sepsis Guideline: An Indication for Precise Medicine?

The newly released 2021 sepsis guideline integrated the most updated evidence to make recommendations [1]. Many recommendation items were added as compared to the previous sepsis guidelines [2]. However, it is obvious that there is still a lack of evidence to support a strong/certain recommendation to be made for most clinical questions. There is large proportion of low quality and uncertain areas in these recommendations [3, 4]. Furthermore, most of the strong recommendations are not based on high-quality evidence but they are based on best practice statement. In best practice statement, the treatment or interventions are usually considered to be unethical if they are implemented or withdrawn, and randomized experimental trials are unlikely to be conducted, such as the prompt initiation of resuscitation after septic shock is confirmed, follow-up for physical, cognitive, and emotional problems after hospital discharge. Furthermore, the authors of the guideline also recommended implementation of individualized treatment in a variety of aspects including the timing of ICU admission, timing of tracheal intubation, sedation and the choice of fluid type (Table 1). However, such recommendations are not clinically implementable, and such individualized treatment can vary substantially across physicians and institutions. There are several reasons why there is lack of high-quality evidence to support clinical decision making in the critical care setting. First, the critical illness encompasses a heterogenous population [5, 6]. Although many diseases can share common clinical phenotypes at their critical stages, the difference in etiology and pathogenesis generate the problem of case-mix of the study population. Take the sepsis as an example, many clinical trials enrolled sepsis patients as the study population, irrespective of the disease stage, site of infection and the immune condition [7–9]. The case-mix of the study population can partly explain the null effect of the study drugs/interventions. For example, the 2021 sepsis guideline makes a weak recommendation for the IV use of hydrocortisone for persistent shock because the desirable effects do not seemly outweigh the undesirable effects. If we classify sepsis into several subphenotypes and then to investigate the potential desirable and undesirable effects in these subgroups, a certain subgroup may display large beneficial effects as compared to other subgroups. In this situation, the recommendations can be made depending on the subphenotypes of sepsis. However, such investigations are still in its infancy and the results requires more verifications with randomized experimental trials [10, 11]. The effectiveness of a classification system requires randomized controlled trials to test, in which the intervention is made depending on the subphenotypes of sepsis, while the control group is treated conventionally without knowledge of the subphenotypes. The usefulness of the classification system can only be formally tested in this way. Second, the heterogeneity of sepsis is not only present at individual level, but at different stages of sepsis. For a given patient, resuscitation strategy should be tailored at different stages. It is well known sepsis patients typically experience several stages including hyper-inflammatory response, immunosuppression and prolonged chronic critical illness associated with persistent inflammatory response. These stages are overlapped and can transition from one to another depending on the treatment strategy. Thus, the * Zhongheng Zhang zh_zhang1984@zju.edu.cn

The newly released 2021 sepsis guideline integrated the most updated evidence to make recommendations [1]. Many recommendation items were added as compared to the previous sepsis guidelines [2]. However, it is obvious that there is still a lack of evidence to support a strong/certain recommendation to be made for most clinical questions. There is large proportion of low quality and uncertain areas in these recommendations [3,4]. Furthermore, most of the strong recommendations are not based on high-quality evidence but they are based on best practice statement. In best practice statement, the treatment or interventions are usually considered to be unethical if they are implemented or withdrawn, and randomized experimental trials are unlikely to be conducted, such as the prompt initiation of resuscitation after septic shock is confirmed, follow-up for physical, cognitive, and emotional problems after hospital discharge. Furthermore, the authors of the guideline also recommended implementation of individualized treatment in a variety of aspects including the timing of ICU admission, timing of tracheal intubation, sedation and the choice of fluid type (Table 1). However, such recommendations are not clinically implementable, and such individualized treatment can vary substantially across physicians and institutions. There are several reasons why there is lack of high-quality evidence to support clinical decision making in the critical care setting.
First, the critical illness encompasses a heterogenous population [5,6]. Although many diseases can share common clinical phenotypes at their critical stages, the difference in etiology and pathogenesis generate the problem of case-mix of the study population. Take the sepsis as an example, many clinical trials enrolled sepsis patients as the study population, irrespective of the disease stage, site of infection and the immune condition [7][8][9]. The case-mix of the study population can partly explain the null effect of the study drugs/interventions. For example, the 2021 sepsis guideline makes a weak recommendation for the IV use of hydrocortisone for persistent shock because the desirable effects do not seemly outweigh the undesirable effects. If we classify sepsis into several subphenotypes and then to investigate the potential desirable and undesirable effects in these subgroups, a certain subgroup may display large beneficial effects as compared to other subgroups. In this situation, the recommendations can be made depending on the subphenotypes of sepsis. However, such investigations are still in its infancy and the results requires more verifications with randomized experimental trials [10,11]. The effectiveness of a classification system requires randomized controlled trials to test, in which the intervention is made depending on the subphenotypes of sepsis, while the control group is treated conventionally without knowledge of the subphenotypes. The usefulness of the classification system can only be formally tested in this way.
Second, the heterogeneity of sepsis is not only present at individual level, but at different stages of sepsis. For a given patient, resuscitation strategy should be tailored at different stages. It is well known sepsis patients typically experience several stages including hyper-inflammatory response, immunosuppression and prolonged chronic critical illness associated with persistent inflammatory response. These stages are overlapped and can transition from one to another depending on the treatment strategy. Thus, the treatment strategy should also be tailored in response to the state transition. Such a clinical question can be formalized by the idea of reinforcement learning which is borrowed from computer science [12]. Reinforcement learning is one of three basic machine learning paradigms, alongside with the supervised learning and unsupervised learning. In the reinforcement learning framework, the intelligent agents ought to take actions in an environment in order to maximize the notion of cumulative reward [13]. Such learning algorithm has been successfully implemented in the management of sepsis in the literature [12,14], but its applications to real clinical practice is still in its premature stage. Alternatively, the sequential decision rules for sepsis resuscitation can also be formalized by using the dynamic treatment regime (DTR) model. Treatment choices made for a particular patient under a dynamic regime are based on that individual's characteristics and history, with the goal of optimizing his or her longterm clinical outcome [15]. A dynamic treatment regime is analogous to a policy in the field of reinforcement learning. DTR model has been reported to solve the resuscitation problem in septic shock or mechanical ventilation in acute respiratory failure [6,16].
Third, critical illness included a variety of populations and there are much more clinical questions than the number of trials that can be conducted. The management of sepsis, for instance, requires multidisciplinary teams from resuscitation intervention to rehabilitation. In the initial stage of septic shock, surgical operations might be required to remove abscess or repair visceral organ perforation. At the recovery stage, the rehabilitation team is indispensable to accelerate functional recovery. Thus, numerous clinical questions are raised in drafting the guideline. Currently, most clinical trials involving sepsis are mainly focused on the initial resuscitation phase, while less is known about the post-resuscitation or rehabilitation stage. In the 2021 sepsis guideline [1], the clinical managements involving follow-up or post-discharge care are usually based on best practice recommendation, indicating a lack of evidence in this overlooked area.
In conclusion, while there are tremendous advances in the management of sepsis in the past few years, there is still a large room to improve the quality of evidence. Conventional research paradigm that encompasses a heterogenous population in a single trial may not be suitable for the clinical syndrome sepsis. The concept of precise medicine should be implemented in the trial design and clinical management of sepsis [17]. The advances in big data and machine learning can be exploited to implement individualized treatment strategy in the critical care setting, beyond the sepsis syndrome. Sepsis is a good prototype model that can be used to test the feasibility that artificial intelligence may help to improve patient care in the ICU.
Author Contributions ZZ and LC conceived the idea and drafted the manuscript; HC critically reviewed the manuscript; YH provided profound insights into the study work.

Conflict of interest There is no conflict of interests.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long Table 1 Recommendations in support of personalized management of sepsis We recommend individualizing the timing of ICU admission. It should ideally be within minutes in severely ill patients but can be less urgent in less severe cases. No time limit is applicable for all patients. The decision may be influenced by the level of care available within ward areas and by ICU bed availability and, of course, by the physiological status and reserve of the patient We recommend individualizing the need for and timing of tracheal intubation, based on careful clinical assessment, including level of consciousness, respiratory rate and work of breathing, hemodynamic status, and assessment of gas exchange. Delaying tracheal intubation may lead to respiratory and even cardiac arrest, with dire consequences, yet premature use of invasive mechanical ventilation can expose the patient to ventilator-induced lung injury, distant organ complications, and increased risk of nosocomial lung infection We recommend individualizing sedation therapies, recognizing that many septic patients need little or even no sedation. Tracheal intubation per se is not a sufficient indication for administration of sedative agents. Sedative agents reduce vascular tone and myocardial contractility, and may also alter immune function We recommend individualizing the type of intravenous fluid administered. For example, albumin administration may be considered in an edematous patient with profound hypoalbuminemia or prolonged non-response to crystalloids We recommend individualizing administration of inotropic agents when tissue hypoperfusion relates to impaired cardiac function (documented at least by echocardiography). The choice and the dose of the inotropic agent should be based on individual hemodynamic monitoring with repeated measurements We recommend involving senior colleagues and consultants, especially since guidelines are most useful for non-experts. Team work, communication and multidisciplinary teams are essential aspects. One of the most overarching recommendations is to seek for guidance from other colleagues and to clearly document the rationale for an intervention-be it recommended or not in the guidelines as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.