Applying artificial intelligence technology to support decision-making in nursing: A case study in Taiwan

Nursing staff and nursing practice

Nurses are on the front line of clinical duties and serve as an important safeguard for patients’ safety. In the current arrangement of nursing staff shifts in Taiwan, one nurse, on average, takes care of 7–12 patients on a day shift and 12–30 patients on a night shift. ² In Taiwan, only 357.9 nurses are available for every 100,000 people, a number significantly lower than 771 in New Zealand, 724 in the United Kingdom, 492.1 in Singapore, 782 in the United States, and 763.8 in Japan.^6,7 As a result, the nurses in Taiwan are overworked, and many quit their jobs, mostly due to stress. According to previous research, junior nursing staff members who quit their jobs accounted for 20–32 percent of the total number of nurses who resigned, much higher than the average 9 percent in other professions. ⁸ The following are the three major reasons why junior nurses quit: (1) lack of clinical experiences and nursing skills, (2) poor coordination with colleagues, and (3) work overload affecting daily life. ⁹ Studies indicate that new clinical nurses face multiple forms of stress due to both internal and external factors. They thus tend to make mistakes. However, these mistakes are often caused by flawed systems rather than individual negligence. ¹⁰ In an intricate medical system, the mistakes could be linked to all kinds of factors. ¹¹ However, the key factors are often associated with the design of the entire workflow. As a poorly executed process could result in excess pressure on the nursing staff, the development of an efficient nursing information system to support nurses is of vital importance. ¹²

The first stage in the clinical healthcare process is the nursing assessment—assessing the patient’s vital signs, recording the patient’s major complaints, and conducting a physical examination. The commonly used Gordon’s 11 functional health patterns in clinical nursing care collect subjective and objective data used in the second stage, which includes an analysis of the gathered data and the formulation of a nursing diagnosis. In the third stage, preliminary diagnoses are used to set nursing care targets. The fourth stage involves scheduling and applying treatment methods in accordance with the targets set in the third stage. Finally, in the fifth stage, healthcare providers evaluate whether the targets have been achieved and whether the problems have been resolved. Afterward, they determine the need to continue nursing care or to reevaluate patients to validate the process flow or to detect any new problems. All together, these various stages make up the nursing process. ¹³

Decision-making in nursing

A nursing decision is defined as the determination of health problems, priorities, and separate nursing practices by a nursing staff member who combines his or her comprehensive physiological, psychological, spiritual, and social views of a nursing case with his or her own knowledge, past experience, and instinct before making a judgment. ¹² From several recent studies, we have found that the actual nursing issues faced and nursing practices offered by the nursing staff exceed the items disclosed in most nursing records. In the current medical environment, one in which patients develop quick and unpredictable symptoms, medical techniques and equipment are constantly advancing, and the manpower of nursing staffs is continuously shrinking, it is not easy to make correct nursing decisions and keep accurate records in a limited amount of time.^13,14 In clinical practice, a nursing staff is always faced with situations that need prudent decision-making. The nature of the decisions to be made lies in making choices, which could be complicated.^14,15 Nurses need to arrive at an effective clinical decision through extensive sources of knowledge and reliable information in a supportive environment. ¹⁶ From research studies and first-hand clinical experience, we have found that when a patient is hospitalized, a nursing staff obtains the patient’s preliminary information by using Gordon’s 11 functional health patterns. ¹⁷ However, due to the shortage of nurses, poor judgment, and a shortened recording procedure, most nursing staff members experience a lot of stress when making decisions. They find it hard to create accurate and complete nursing diagnoses based on the data they collect and analyze in such a short period of time. ¹⁸ Moreover, the recording stage can take substantial time out of a nurse’s busy schedule and is more likely to be interrupted by other duties. Constraints could also affect data gathering and thus result in illegible records and missing information, which could make it difficult for nurses to assess the actual health issues of patients.

A decision support system is a type of information system that helps individuals to make relevant decisions by leveraging databases and the computing capability of artificial intelligence to facilitate decision-making processes.^19,20 Nursing staff members usually have to obtain information and make prompt decisions under time constraints, so they often fail to completely understand the health issues of patients and thus limit their capability to make accurate decisions.^21,22 According to some academic research reports, decision support systems could provide medical staffs with quick and appropriate suggestions for various clinical symptoms and exceptional testing results. These systems could bring more useful information to healthcare providers so that they can make better decisions.^1,3,4,23,24 The main function of decision support systems is to support users and provide relevant information. The results generated by the systems are not necessarily in a particular format but may vary with user inputs. The system database is established initially based on the users’ professional knowledge and experiences. ⁵

Artificial intelligence technology

Nursing informatics has been developed over recent years as IT has become prevalent in applications to nursing practices. Its purpose is to create a nursing-practice-oriented recording model that is patient-centered. The model is expected to develop empirical nursing information through the support of artificial intelligence technology and to facilitate nursing care in order to establish clinical nursing guidance for various kinds of patients. ²⁰ Thus, if its algorithm can be effectively applied to nursing practices involving patient care and for judging work priorities, it could be of great help to healthcare providers. ²¹ Artificial intelligence deals with the exploration, study, and design of machines and equipment that simulate human behaviors. These machines and equipment demonstrate human or near-human intelligence and can think, operate, and use their intelligence functions much as people do. As the serial processes of artificial intelligence are similar to the way in which human neurons function by delivering and interpreting messages, machines using artificial intelligence are also called artificial neural networks. ²⁵ At present, the most popular and representative learning model is a back-propagation neural network (BPN), in which training data are input during learning sessions so that the weights in the network can be incessantly adjusted during the learning process. After the learning process is finished, each input of data into the network can bring immediate effective responses.^26,27

Previously, BPNs were often used to analyze and determine commercial cases. In recent years, the techniques of artificial neural networks have been widely applied in medical decision-making processes. According to research findings, the application of BPNs could enhance the accuracy and consistency in determining the triage and acuity scale in the emergency department. ²⁰ In forecasting medical caring models of discharge planning services, the use of artificial intelligence techniques could help doctors to appropriately identify the appropriate care facilities for patients after their discharge. ²⁸ A research study developed an online enquiry and diagnosis system that adopted an adaptive-network-based fuzzy inference system (ANFIS) to test the classification of diabetes and metabolic syndromes. It further determined patients’ risk of getting diabetes and its severity according to their biochemical test results and lifestyles. This system proved to be valuable for the medical staff, assisting them in making better clinical decisions as well as in discovering and treating diseases preemptively.^7,29

Study process and framework

In this study, we constructed a predictive model based on the Cross-Industry Standard Process for Data Mining (CRISP-DM). The model construction involved the following stages: (1) The major data resources used were charts from patient hospitalizations in the same medical specialty over the last year. (2) The data used were then preprocessed. Preprocessing included the screening, deletion, and regularization of information to reduce the effects of noise and incomplete data that might affect the accuracy of the prediction model developed by this study. (3) Factors that might affect nursing diagnoses were investigated based on relevant literature, which was followed by selection of primary discriminating factors from patient files to be integrated as input variables. (4) The K-fold cross-validation technique was applied to group samples into K subset clusters. K-1 groups were used as training groups and the remaining group was used as a testing group to test validity. Then another group was chosen as a testing group with the remaining K-1 groups as training groups. Mean accuracy of the analysis model was averaged after 10 rounds of testing. (5) Prediction models were constructed and the best model from various trial models was selected through comparisons of model accuracy (Figure 1).

OPEN IN VIEWER

Figure 1.

System framework.

Data collection

We first reviewed 216 medical records from 105 patients discharged from the hospital, and we used Gordon’s 11 functional health patterns (Figure 2) to make a comparison with the system according to the established nursing diagnoses prepared by clinical nursing staff members from June 2011 to March 2012. Senior nurses with at least 4 years of clinical experience were qualified to serve in the main assessor roles. With this sample size, analysis results were more reliable and had reference values. Data from each patient were keyed into SPSS software with 26 fields including coding, age, gender, symptoms (a total of 18 fields), and nursing diagnosis (a total of 5 fields). Data from all the fields were keyed into SPSS software for further analysis.

OPEN IN VIEWER

Figure 2.

Diagram of Gordon’s 11 functional health patterns with various nursing diagnoses.

Experiment

In this study, we used IBM SPSS, the decision tree in Clementine 12.0, a BPN, and the ANFIS in MATLAB as the analysis tools for data mining and statistical analysis. We first used the decision tree and clustering algorithm to find important factors and association rules that could be used for nursing differential diagnosis. We divided the test samples into two groups. One group served as a training sample and the other as the test sample to test the accuracy of the rules. We further compared the prediction accuracy of the previous methods with the ANFIS and the BPN.

Decision tree

First, SPSS was used to import raw clinical data to Clementine for computing the distribution of each variable of the raw data. For the factor selection, principle component analysis was used to filter out unimportant properties. The decision tree of the data mining techniques was used to construct the classification model. We obtained the error rate from the tree diagram and decision tree by generalizing five common nursing problems and 18 factors (symptoms and signs) related to those problems out of the 216 nursing diagnoses in the training and testing groups and bringing them into the decision tree algorithm. The total error rate for the decision tree algorithm was 12 percent, indicating that the accuracy rate of recognition was as high as 88 percent (Table 1).

OPEN IN VIEWER

Table 1.

Rules for determining nursing diagnoses via decision tree.

Rule	Term	Nursing diagnosis	Accuracy	Accuracy %
1	IF Chest tiredness > 0	Low cardiac output	61/61	100%
2	IF Chest tiredness ≤ 0	Risk for fall accident	58/68	85%
	AND four limbs weakness > 0
3	IF Chest tiredness ≤ 0	Risk for insufficiency of perfusion	36/41	88%
	AND four limbs weakness ≤ 0
	AND Poikilothermic > 0

ANFIS

In this study, we used the equation rule in the ANFIS to predict the accuracy of the next nursing diagnosis (Figure 3). MATLAB codes were written to analyze the collected data. ANFIS is an extension of the fuzzy inference system (FIS). ANFIS combines neural network principles and fussy inference principles to construct a network inference model that is a supervised FIS. ANFIS is useful in adjusting optimal parameters to achieve the best predictive effects. As we have already seen, fuzzy systems present particular problems to developers. The if-then rules have to be determined somehow. This is usually done through “knowledge acquisition” from experts, which is a time-consuming process that is fraught with problems (Figure 4; equation (1)). A two-rule Sugeno of ANFIS has rules of the form

If x is A 1 and y is B 1 Then f 1 = p 1 x + q 1 y + r 1

If x is A 2 and y is B 2 Then f 2 = p 2 x + q 2 y + r 2

(1)

OPEN IN VIEWER

Figure 3.

Probability diagram of the nursing diagnosis.

OPEN IN VIEWER

Figure 4.

ANFIS architecture for a two-rule Sugeno system.

Clinical trials

This study conducted clinical testing at two healthcare units in one teaching hospital in northern Taipei. The nursing staffs received a 3-month-long training session before the implementation of the system. Then, the preliminary experience was explored and shared 3 months after the implementation.

Using the English version of SPSS 20.0, we analyzed the basic data to identify the differences in the questions and in the nurses’ satisfaction levels before and after using the nursing information system. We also conducted paired t-tests to check whether significant differences exist. A p value less than 0.05 was defined as statistically significant. Moreover, we developed a questionnaire as part of our research method. The questionnaire included two major parts. The first part was concerned with basic information, and the second was about question establishment and satisfaction scoring. The basic information questions included items on age, education, seniority, and basic computer skills (Word, Excel, and PowerPoint). To confirm the questions’ validity, we compared the questions raised by the nursing staffs and the ones suggested by the systems, and then the nurse leaders assessed their accuracy. For satisfaction scoring, the survey adopted a 5-point Likert scale (1 = very unsatisfactory, 2 = unsatisfactory, 3 = near satisfactory, 4 = satisfactory, and 5 = very satisfactory). Among the three units tested using the questionnaire, a total of 36 nurses have used such a system. The statistical result showed a reliability coefficient α of 0.83.

Descriptive analysis

For the 105 cases collected, most of the patients were males with an average age of 51 ± 2.53 years. Among the 11 functional health patterns, health perception, nutrition, activity, and sleep–rest were found to be statistically significant influencing factors. Among nursing diagnoses, acute pain, discomfort, and activity intolerance were most common. The average age of the 36 nursing staff members was 28.92 ± 4.02 years, and the average work experience was 5.95 ± 1.8 years. Half of the nurses were university educated and the other half were college educated. They all have good self-perceptions of their computer operation skills (Table 2).

OPEN IN VIEWER

Table 2.

Basic nursing staff information (N = 36).

Variable item	(n = 36)
	Mean ± SD	Number of people (%)
Age	28.92 ± 4.02
Seniority	5.95 ± 1.8
Education degree
College		18 (50.0)
University		18 (50.0)
Graduate and above		0 (0)
Basic computer skills
Poor		0 (0)
Good		28 (77.7)
Excellent		8 (22.3)

SD: standard deviation.

Comparing different prediction models

We used the BPN testing by SPSS Clementine 12.0 to determine the hidden number of neurons and the learning ratio α. The BPN included only one hidden layer, and the input layer included 18 neurons. Therefore, the numbers of neurons in the hidden layer were set at three, five, seven, and eight for running the tests. Finally, the output layer of the network included five neurons with a final accuracy rate of 92.41 percent and an error rate of 7.59 percent (Table 3).

OPEN IN VIEWER

Table 3.

Parameter setting in the artificial neural network.

Parameter	Value
Mining capability	Back-propagation neural network
Data input	113 cases in the training group
Accuracy rate	92.41%
Error rate	7.59%
Input layer	11 neurons
Hidden layer	4 neurons
Output layer	5 neurons

We generalized 18 possible factors related to 5 common nursing diagnoses out of 108 cases and then put the 216 nursing diagnoses into training and testing groups. The ANFIS model in MATLAB was used to verify the accuracy of the clinical diagnoses. The ANFIS adopted in this research yielded a final accuracy rate of 84.82 percent and an error rate of 15.18 percent.

Effectiveness of artificial intelligence technology

We selected only the first three targets generated by the BPN as our first choices in screening the problems of nursing diagnoses. When consolidating the 18 related factors that could affect the nursing diagnoses, we created a forecasting model of the BPN for the 123 sets of training data by using the Clementine software and then used 93 sets of patient data for testing purposes. After using Clementine to determine the accuracy of the aforementioned training data, we achieved an 87.41 percent accuracy rate in predicting patients’ nursing diagnoses.

This nursing information system can enhance the capability of nurses to provide accurate nursing diagnoses. We conducted a clinical test in one teaching hospital in northern Taipei by collecting 54 medical records that were reviewed by senior nurses. The percentage of nursing diagnoses suggested by the information system that coincide with those made by the nursing staff over the total samples is as much as 87 percent. The percentage of the nursing diagnoses made by both the system and the nursing staff that are concurrently in the top three possible results over the 54 samples is 74 percent.

The nursing information system could also improve the job satisfaction of nurses. Results showed that overall patient healthcare was improved, work satisfaction concerning nursing diagnoses was enhanced from 41.1 percent to 75 percent, and the time spent on decision-making was reduced from 35.5 to 19.8 min (Table 4).

OPEN IN VIEWER

Table 4

Satisfaction and time spent for the forecasting method.

	(n = 36)
	Satisfaction	Time	t	p
Before implementation	41.1		17.251	0.001*
After implementation	75.0		40.599	0.001*
Before time		35.5	11.995	0.001*
After time		19.8	11.223	0.001*

*

P value < 0.05.