A model-based clinical biomarker for sepsis diagnosis in critical care patients.
Thesis DisciplineMechanical Engineering
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
Sepsis, severe sepsis, and septic shock are stages of a medical emergency characterised by an intensifying whole-body immune response to infection leading to organ dysfunction, shock, and ultimately death. Importantly, these stages do not represent an intensifying infection, but rather the body’s intensifying immune response to infection. Yet, despite advances in modern critical care medicine, sepsis remains common, increasingly costly, and often deadly. Time to initiation of effective antimicrobial therapy following sepsis-induced hypotension (i.e. septic shock) is the single strongest predictor of outcome over any form of treatment. Furthermore, early treatment reduces sepsis mortality. Importantly, a challenge in the early identification of sepsis is that infection is not always clinically evident. Gold standard blood culture microbiological results return only in retrospect with significant delay. Additionally, there are no biochemical and immunological biomarkers with sufficient performance for routine use in critical care. Finally, protocolised categorisation using the ACCP/SCCM sepsis definitions in real-time is erratic and often reflects misclassification, heterogeneous categorisation, and exclusion. Thus, there remains a serious need for early, accurate, time-dependent, patient specific diagnostics for sepsis available at the bedside in real-time for clinical decision support. Mathematical models of physiology developed from clinical data can identify patient-specific parameters, in particular, model-based insulin sensitivity (SI), which is related to patient condition and sepsis state. A multivariate biomarker has been shown to link model-based SI and clinical measures to septic shock. This thesis further develops a model-based sepsis diagnostic for severe sepsis from model-based SI , temperature, heart rate, respiratory rate, blood pressure, and SIRS score. Study data was obtained from patient records of 36 adult sepsis patients in the Christchurch Hospital ICU, where the ACCP/SCCM sepsis definitions were used to categorise hourly sepsis state, resulting in 213 hours of severe sepsis and septic shock cases and 5858 hours of SIRS and sepsis controls. Kernel density estimates (KDE) using the Bayes classifier were used to estimate class conditional joint probability density profiles of the clinical predictors and for classification. The unknown patient hour to be classified was tested against these established datasets, with the result being a classification into either the case or control group. The classifier performed with the greatest stability and accuracy when using the product kernel, 0.5 prior probabilities, and Cholesky transformation. Optimal diagnostic performance from the receiver operating characteristic (ROC) curve was determined as 0.78 (0.69–0.94) sensitivity, 0.83 (0.76–0.94) specificity, 0.87 (0.78–0.99) AUC, 0.10–0.36 PPV, 0.99–1.00 NPV, 4.48 (2.88–15.70) LHR+, 0.27 (0.06–0.41) LHR-, and 16.83 (7.04–262) DOR at an optimal posterior probability cutoff value of 0.31. Thus, kernel implementation of the Bayes classifier given bedside clinical measurements can provide a useful posterior probability for clinical decision making in real-time. An independent classifier was developed whereby the ACCP/SCCM classification criteria were independently evaluated and summed, providing a 25.8% disease prevalence (1690 of 6550 hours). Similarly, the KDE estimation and classification method was used, resulting in optimal diagnostic performance of 0.86 (0.81–0.94) sensitivity, 0.85 (0.79–0.95) specificity, 0.92 (0.88–0.99) AUC, 6 (4–18) LHR+, 0.17 (0.06–0.24) LHR-, 0.57–0.86) PPV, 0.92–0.98) NPV, and 34 (16–300) DOR at an optimal posterior probability cutoff value of 0.49. The diagnostic results show high accuracy as a potential severe sepsis diagnostic and monitoring response to sepsis interventions in real-time. Thus, relaxation of the hierarchical and concurrent criteria in the ACCP/SCCM definitions captured the more staged and clinically observed evolution of sepsis over time, including plateaus of septic shock treatment during administration of IV fluid resuscitation. Therefore, it is an improved, objective metric especially for real-time diagnosis and monitoring of response to disease and treatment. A hidden Markov model (HMM) was developed to link observed clinical measurements to unobserved sepsis states and to include time-dependency. A HMM topology was defined to represent the study variable relationships, given the observed time series of physiological variables. In particular, the topology defines transitions for the hidden states and the distributions of the observations conditioned on each hidden state. Thus, the labelled data can be used to estimate the transition probabilities of the hidden sepsis states. The conditional distributions, P (observation–sepsis state), were found using the joint probability densities using kernel density estimates. Finally, the hidden states were estimated by determining the most probable path of the joint probability of the observed sequence and the hidden sequence. Upon determining the posterior probability of a patient sepsis state, the patient hour is compared against the established dataset and diagnostic performance from the ROC curve was determined for resubstitution, repeated holdout estimate, and leave one out estimate. The HMM performed with 0.59–0.95 sensitivity, 0.61–0.96 specificity, 1.54–23.96 LHR+, 0.05–0.66 LHR-, 0.63–0.99 AUC, and 2–474 DOR. The state transition probabilities were shown to be independent of sepsis categorisation definitions. Furthermore, the observed clinical signs are linked to hidden sepsis state, yet are most accurate when the model is trained on the patient data. Thus, the HMM has the most potential as a real-time, patient-specific model to reduce the variability of diagnosis due to inter- and intra-patient variability. Overall, this thesis develops and characterises a range of model-based metabolic biomarker linked sepsis diagnostics. The analysis of their efficacy is taken to a statistically valid level not typically seen in the medical literature and provides significant new insight into how diagnosing sepsis is affected by prevalence and lack of clarity in the specific criteria used. The diagnostics created are all novel for their real-time, hour-to-hour approach compared to the typical multi-hour or daily evaluation typically used that provides detection only with significant delay. Thus, the approach itself offers new potential. The sum of this work provides a significant step forward and clear foundation from which to develop objective, automated, real-time sepsis diagnostics, a prototype for clinical validation, as well as providing significant new insight into sepsis, its diagnosis and how it is viewed clinically.