Role of artificial neural networks in prediction of survival of burn patients—a new approach

Abstract

A burn patient may require the most complicated treatment regimes encountered among trauma victims. Predicting the outcome of such treatment depends on several factors which have non-linear relationships. Traditional methods in prediction are “logistic regression” and “maximum likelihood”. In this study, an artificial neural network (ANN) is used for computing survival among burn patients admitted to the “Motahary Burn Center”, during a 1 year period (1996–1997). Fifteen different observations, such as total body surface area (TBSA), rescue time, admission period, surgery, inhalation injuries, etc. were obtained, retrospectively. A normal feed forward ANN was developed by Thinkspro software. It has 15 input-units, two hidden layers, and one output-unit. Survival was higher in males, those in whom early fluid resuscitation had been initiated and in patients in the middle of the age spectrum (P<0.0001). Strong correlations with these factors were noted.

In the training phase, the ANNs accuracy reached 90%. In this study, the ANN has been applied for the first time to predict burn victim survival. This study can enable a different view point to help burn center physicians in the prediction of survival of their patients.

Introduction

A burn injury is a disastrous trauma and can have very wide ranging complications making such cases difficult to manage and treat. Severe morbidity in victims is observed and mortality rates may be high. Burn patients generally have a long period of stay in the hospital, approximately 1.5 days per total body surface area (TBSA).

Outcome may be diverse. Some patients may have healing of wounds and develop minimal scars. Others may develop severe contractures, hypertrophic scars and various deformities. Some unfortunate victims may have suffered inhalation injury, severe chemical or electrical injuries. Victims may develop wound infections and septicaemia and eventually die as a result of these complications.

Several factors influence outcome among patients including age, TBSA, presence of inhalation injury, previous medical conditions, infections, pneumonia and septicaemia, depth of burn wounds and so on.

At the first examination the physician is usually confronted with the commonly encountered questions:

“Can the patient overcome this severe condition?”

“Will he live?”

“What is the percentage of his prognosis?”

Those questions are very difficult, challenging to the intellect and time consuming for the physician in the evaluation of the patient and the assessment of the probability of survival. Some linear statistical methods such as linear regression may help the physician in some aspects of problem, but evaluation of the whole condition of patient is generally not within the capability of these statistical methods.

Computer software and programs can help the physician to address the problem in three different ways:

First: linear connection software (Fig. 1), this is characterized by simple forward connection and/or simple “if … then phrases” can answer very simple questions with only “yes–no” type answers.

Second: expert systems (Fig. 2), these systems can save numerous phrases and consider several conditions or “if … then phrases” and may consider positive and negative influences. In this way answers to some complex questions can be found. These systems obviously need several previously defined phrases (or facts) and someone should write or input these facts using a special computer language into the system. The writer must consider all situations which may occur in the processing of the question and also consider very rare conditions (this is called supervised training). This means that these systems are very time consuming for development time before compiling and running. Some conditions may be omitted or neglected. This will reduce the reliability of the system.

Third: intelligent systems (Fig. 3), these systems simulate processes and systems in the human brain and neuronal connections, and can learn and judge. These systems have several units or neurons which resemble the neural cells in brain, and are called “artificial neural network (ANN)”. Like the brain, ANN can recognize patterns, manage data and most significantly “learn” in a way which simulates human intelligence. ANN constantly improves its functional accuracy, it does not need formulated factual input. These systems are used to supplement and even replace experts. They can classify and interpret various forms of medical data. Since predictive ability is a major advantage of ANN, such systems have the potential to assist with description of morbidity and help with mortality.

In this study, in the first instance retrospectively data was extracted retrospectively (including outcome) from all of burn patients files who were admitted during 1 year (1996–1997) and interpreted with traditional linear statistics such as logistic regression, Kaplan–Meier, and Cox proportional hazard in order to calculate mortality rate of patients.

In a second stage, an ANN system was constructed and patients were divided into two groups. One group served as a training group and the other as test group to evaluate total accuracy and sensitivity of the network.