The development and application of novel intelligent scoring systems in critical illness

Sim, Malcolm A.B. (2015) The development and application of novel intelligent scoring systems in critical illness. MD thesis, University of Glasgow.

Abstract

Scoring systems in medicine are not a new concept. There are examples from the early 1950s, from around the same time as the polio epidemic in Copenhagen resulted in the birth of modern Intensive Care. Many scores have subsequently been developed specifically for Intensive Care patients. The majority summarise the overall physiological state of the patient in a variety of different ways.

A clinical interest in ascertaining whether haemodialysis causes cardiovascular instability in Intensive Care patients led to an initial simple experiment examining stability using a small number of cardiovascular parameters. It became apparent that to answer the question properly a physiologically based score which could be calculated automatically in real time, and which took into account the level of physiological or pharmacological support the patient was receiving would have to be developed, to counter or to mitigate the drawbacks of the main scoring systems in common use at the time.

This thesis describes the development and first stage in the validation of a novel physiologically based scoring system for Intensive Care patients which overcomes some of the major disadvantages of existing scores. The score was then used to investigate other clinical questions. Myocardial damage in Intensive Care is common and associated with a poor outcome. Aspects of the developed score were used to ascertain if it is possible to detect and predict myocardial damage occurring in Intensive Care patients based on physiological disturbance rather than a rise in biomarkers. The score was subsequently used to examine Intensive Care patient outcomes.

The introductory chapter describes the history of Intensive Care, the mechanism of data collection for patients in Scottish Intensive Care Units and its analysis to enable comparison of different units. Reviewing currently available scoring systems places this work in context and highlights the need for a new score. An overview of renal replacement therapy modalities follows, as an interest in cardiovascular stability during haemodialysis led to the idea for a new scoring system. Myocardial damage in Intensive Care patients is common and indicative of poorer outcomes. This is reviewed, as the developed score was used to detect and then predict where myocardial damage was occurring in critically ill patients, based on physiological disturbance rather than on raised biomarkers.

In Chapter 2, data from dialysis sessions in critically ill patients was collected, prc-processed, and analysed for cardiovascular instability. Using an arbitrary definition of instability as a 20% change in mean arterial pressure or heart rate in either direction, 65% of dialysis sessions were stable and 35% unstable. This simple experiment suggested that haemodialysis is less cardiovascularly destabilising than previously believed. However a major deficiency was the lack of consideration of the level of physiological support required during dialysis. To investigate this and other clinical problems better, it became apparent that a new score would have to be developed.

Chapter 3 describes the development of a novel quantitative score which takes into account the amount of physiological and pharmacological support a patient is receiving. Physiological parameters were separated into those recorded regularly and those recorded intermittently. They were subsequently divided into ranges, scoring increasing points depending upon the degree of derangement. Ranges were based on an extensive literature search, currently available scores, and clinical opinion. Two key parameters viz. mean arterial pressure and oxygen saturation, were then weighted against a range of factors which can either increase or decrease their value. A score of instability could then be calculated by adding points for the weighted and unweighted parameters. After reflection using common clinical scenarios, some of the points scored in different ranges and weightings were revised to give the final quantitative score.

In Chapter 4, the quantitative score was tested against data sets from actual Intensive Care patients to produce graphs of overall cardiovascular stability against time. Although this approach did capture improvements and deteriorations it had several disadvantages. It captured the expertise of a single clinician only, gave an arbitrary number which could be difficult to interpret, and the emphasis given by the clinician to the relative importance of different physiological or pharmacological parameters would not be obvious to others. Clinical reflection led to a new approach to the problem, viz. the development of the 5 point qualitative scale described in Chapter 5.

Chapter 5 describes the development of a 5 point qualitative score for cardiovascular instability, underpinned by complex physiological rules, and capturing the expertise of several senior Intensive Care Clinicians. This is the Intensive Care Unit - Patient Scoring System (ICU-PSS). I scored data sets comprising thousands of predominantly hourly commonly recorded physiological and pharmacological parameters on a 5 point scale of cardiovascular stability (A to E). I also described rules in the form of different parameter ranges to indicate why I had scored time points as stable (A) through to unstable (E). These rules were incorporated into a computer programme which scored unseen data sets which I also then scored. The computer’s predicted A to E score based on these rules and my own score were compared in a confusion matrix. Mismatches with the computer prediction (based on my initial rules) were analysed and I either rescored the data if I considered that I had not assigned the correct level of instability, or modified the rule base. Through this process clinical expertise was better captured. This process was repeated with two other clinicians using my rules as a starting point. This led to further refinements of the rule base. The result was a sophisticated set of rules underpinning a 5 point, easily understandable scale of cardiovascular stability crystallising the expertise of 3 senior Intensive Care clinicians.

The ICU-PSS was tested in a discrimination experiment to ascertain if clinicians could agree with the score moving in a one step and two step change. This is the first stage in full validation of the score

In Chapter 6, the first stage in the validation of the ICU-PSS is described, using 10 clinicians from a city teaching and a district general hospital. It was hypothesised that if they were shown two consecutive hourly time points of physiological data from real patients and asked whether they were improving or deteriorating, they should agree with the ICU-PSS score in more than 50% of cases (random chance). In two discrimination experiments the consultants were, in random order, shown 4 examples of each type of two step improvement or deterioration in the score, e.g. A to C, and 4 examples of each type of one step change, e.g. E to D. In the two step experiment there was 92.9% agreement with the score, and in the one step change experiment, 90.9% agreement. Both were highly statistically significant.

Chapter 7 describes the first of the applications of the validated score. Myocardial damage is common in Intensive Care patients and is an independent risk factor for both short and long term mortality. The mechanism in Intensive Care patients is likely to be the so-called type II damage caused by extremes of physiological derangement leading to a myocardial oxygen supply and demand imbalance. I hypothesised that it should be possible to use aspects of the score to confirm and subsequently predict where this damage is occurs based on physiological disturbance alone rather than on a rise in cardiac biomarkers. Two clinicians agreed that a subset of the level E, D and C rules from the ICU-PSS occurring in 3 out of 5 consecutive time points would represent conditions likely to lead to myocardial damage in the critically ill. Data sets with known sequences of troponin rises were scanned to ascertain if the above conditions were met around the time of a troponin rise within a sequence of troponin rises, given the natural decay of troponin. This was indeed the case in 75.8% of cases (95% CI: 57.7% to 88.9%). Similarly this set of conditions was applied to the same data sets, looking at time periods before a first troponin rise. These conditions were met in 87.5% of cases (95% CI: 61.6% to 98.1%). However, as the confidence intervals are wide (and also for the positive and negative predictive values of these tests), this early work is at best hypothesis generating. It will have to be repeated using much larger data sets.

In Chapter 8, the correlation between the mean ICU-PSS score and outcome was examined. A data set of patients was prepared from Ward Watcher with an approximate 50:50 split of medical and surgical diagnoses. The physiological data from these patients was extracted from CareVue and anonymised. A mean ICU-PSS score was calculated for different points during the patient stay. The data were analysed to ascertain if there were differences in mean ICU-PSS scores at different time periods among the survivors and non-survivors within the medical and surgical groups. There is a suggestion that the mean scores are different in certain patient groups between survivors and non-survivors. However, at the time this work was undertaken the computing system used was not yet able to apply appropriate statistical tests. Future work will focus address this problem and also examine the different proportions of the patient’s stay spent in different categories of the score. This would avoid the difficulties above of converting ordinal to numerical data.

In a final analysis I ascertained the relationship between degree of any troponin rise and outcome, in the population of patients at Glasgow Royal Infirmary. In a study of 100 consecutive patients, troponin rises were grouped into three categories. These were low (0.04-0.19), medium, (0.2-1.99) and high (≥2.0 micromoles/litre). Intensive Care mortailty was 13.3%, 22.7% and 40% respectively. This association is consistent with findings from similar studies elsewhere in the literature.

In summary, I have developed a quantitative score of cardiovascular stability, and have developed, and partially validated, a more effective qualitative score for use in Intensive Care patients. I believe it overcomes the salient disadvantages of other currently available scores. I have demonstrated that it may be possible to confirm the presence of, and detect, where myocardial damage is occuring. Work thus far suggests that there may be an association between this score alone and outcome. Future work will focus on translating the score into a bedside monitor to give a continuous reading of the overall physiological state of the patient, to detect deterioration before it becomes clinically obvious. Characteristic patterns of deterioration associated with impending myocardial damage will be displayed at the bedside with the prospect of earlier intervention aimed at preventing myocardial damage and its associated poor outcome.

• Item Type: Thesis (MD)
• Qualification Level: Doctoral
• Keywords: Intensive care, scoring systems, haemodialysis, myocardial infarction, outcomes
• Subjects: R Medicine > R Medicine (General)
• Colleges/Schools: College of Medical Veterinary and Life Sciences > School of Medicine, Dentistry & Nursing
• Supervisor's Name: Kinsella, Prof. John
• Date of Award: 2015
• Depositing User: Dr Malcolm A B Sim
• Unique ID: glathesis:2015-6512
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 03 Jul 2015 11:55
• Last Modified: 07 Jul 2015 10:57
• URI: https://theses.gla.ac.uk/id/eprint/6512