3. Marco teórico
3.1. Las revistas científicas
3.1.2. Funciones, características y organización
In successive roles of soldier’s care, clinicians should take several decisions based on the patient’s condition and the available capabilities. Thus, having a model that can capture the rapid progress of the patient’s condition is very important. Different trauma scoring sys- tems have been developed to summarise the patient’s condition (see Section 4.4). The main problems of the existing scoring systems are: (1) the scores cannot be estimated or they can be less reliable when any necessary information is missing, (2) they are static, meaning that they cannot capture the progress of the patient’s condition, (3) important elements in combat trauma care such as: evacuation times, treatments, sentinel events, and blast injuries are missing, (4) more complicated reasoning techniques cannot be performed and (5) they are intended to be used retrospectively.
These limitations demonstrate the need of developing a novel methodology for creating a model that has the following characteristics: (1) capture the rapid patient’s progress, (2)
capture the dynamic process of clinical decision making, and (3) account for all the data characteristics as described in Section 4.5.3.
In Chapter 5, we propose a method for building a CDS BN that captures the progression of an acute condition and gives predictions in successive stages of care. The proposed BN can reason in a manner consistent with clinical knowledge without being limited by the observed data. The combat trauma care described in the chapter is used as a case study.
Mortality and morbidity review meetings have been used as a quality assurance to review combat trauma care. Currently, the review panel judges the outcome from a preventabil- ity and salvageability point of view. Their judgement is subjective, and it is based on the clinical and military notes and the available scoring systems. However, the existing scor- ing systems can only describe the injury severity. As a result, a more objective tool that consider treatments and their effect on soldier’s salvageability could add an additional as- surance to the current practice.
In Chapter 6, we propose an alternative use of the model developed in Chapter 5. Partic- ularly, we explain how we can use the developed model, alongside the current practice, as a healthcare governance tool to answer counterfactual questions about the effect of treat- ment decisions, other than those occurred to assist the DMS mortality and morbidity review meetings.
Chapter 5
Developing a Progressive Bayesian
Network for Modelling the Evolution of
an Acute Medical Condition and the
Dynamics of Clinical Decision Making
Developing models that accurately capture the way clinicians gather information and make decisions during the stages of a patient’s care is challenging. The modelling problem is especially difficult for acute conditions that evolve rapidly over a short timescale. In Sec- tion 2.6, we presented various time-based BNs that have been developed to address this challenge, but they make assumptions that are rarely true in acute medical conditions (see Section 3.2). In this chapter, we present a methodology for developing a progressive BN that models the rapidly evolving progress of an acute condition, and captures the way clini- cians gather information and make decisions in successive stages of care. To overcome the typical lack of data for this type of problem, we use a combination of expert knowledge and available data to produce a causally coherent BN. The method is illustrated throughout using a comprehensive case study on predicting the mortality risk to combat trauma casu- alties in two successive stages of the soldier’s care. In Chapter 4, we described the context of combat trauma care, where the limitations of the existing trauma models and the char- acteristics of the datasets highlight the need for a new modelling technique. The proposed model could be used to support clinical decision making in successive stages of care.
5.1
Introduction
One of the major challenges in developing useful clinical decision support systems for specific medical conditions is to model the dynamic process of decision making as the condition progresses. Unlike ‘chronic’ conditions, such as diabetes, which develop over a long timescale, an ‘acute’ medical condition is one that develops suddenly over a short timescale and in which, symptoms and treatment capabilities can vary greatly and rapidly at each stage of the patient’s care. Consider the example of a person suffering trauma, such as a severe head injury from a car accident. The symptoms, diagnostic tests and treatment strategies vary based on whether the patient is treated in an ambulance, in the ED, or in the ICU. Moreover, the information on which the decisions are based, especially at the begin- ning, might be limited. For instance, in the car accident example, the information at the beginning might be limited to the injury description and physiology. In the ICU, more data about blood results, operation’s response, fluid and drugs given will be available.
Over the years, many researchers have expressed the need to represent the dynamic nature of clinical decision making [5], [146], [152], [258], [10], [3], [120]. As explained in Sec- tion 2.6, many time-based BNs have been proposed. Despite their benefits, modelling an acute medical condition remains challenging (see Section 3.2). In this chapter, we propose a practical methodology for doing so. The methodology captures the way clinicians gather information and make decisions in successive stages of care. Crucially, by exploiting expert knowledge, the methodology is not limited by the available data. The method can gener- ate several predictions in successive stages using a progressive BN. The term ‘progressive’ refers to the evolving non-stationary structure of the model which, at each stage, is an ex- tension of the previous stage. We use the term ‘stage’ and not time slice as we provide predictions in successive stages of the patient’s care. The time interval between each stage is irregular and follows clinicians’ timeline and not data availability. To illustrate and vali- date the methodology we use a clinical case study on predicting the mortality risk to combat casualties in the prehospital setting.
The chapter is organised as follows: in Section 5.2 we briefly introduce the characteristics and the development process of a progressive BN. We illustrate our methodology using a case study, which is described in the same section. The way to define the model’s variables is described in Section 5.3. The process of developing the BN structure is explained in
Section 5.4. In Sections 5.5 and 5.6 we describe the process of parameter learning and elic- itation, and the validation of the model’s performance, respectively. Finally, a discussion is provided in Section 5.7.