Delirium Prediction Using Machine Learning Interpretation Method and Its Incorporation into a Clinical Workflow

Abstract

Delirium in hospitalized patients is a worldwide problem, causing a burden on healthcare professionals and impacting patient prognosis. A machine learning interpretation method (ML interpretation method) presents the results of machine learning predictions and promotes guided decisions. This study focuses on visualizing the predictors of delirium using a ML interpretation method and implementing the analysis results in clinical practice. Retrospective data of 55,389 patients hospitalized in a single acute care center in Japan between December 2017 and February 2022 were collected. Patients were categorized into three analysis populations, according to inclusion and exclusion criteria, to develop delirium prediction models. The predictors were then visualized using Shapley additive explanation (SHAP) and fed back to clinical practice. The machine learning-based prediction of delirium in each population exhibited excellent predictive performance. SHAP was used to visualize the body mass index and albumin levels as critical contributors to delirium prediction. In addition, the cutoff value for age, which was previously unknown, was visualized, and the risk threshold for age was raised. By using the SHAP method, we demonstrated that data-driven decision support is possible using electronic medical record data.

1. Introduction

Learning health systems are self-learning transformation systems in which data accumulated in clinical practice are used to prepare evidence-based medicine [1]. As an example of digital technology’s potential to advance a learning health system, a multicenter study involving 43 hospitals demonstrated a 44% reduction in all-cause central-line-associated bloodstream infections through a universal decolonization strategy in intensive care settings that used real-time patient data [2,3]. However, there are limited examples wherein medical support using real-world data has been structured as part of a hospital improvement activity. In this study, delirium was used as a case study for investigating the efficacy of learning health systems in real-world settings.

Delirium has been reported in 8–17% and 11–51% of emergency and surgical patients, respectively [4]. The occurrence of delirium in elderly patients is associated with mortality, institutionalization, dementia after discharge, and other poor outcomes. Therefore, the management of delirium during hospitalization is crucial [5].

Delirium has been reported globally, and many evaluation tools have been developed to identify patients at high risk of delirium at the time of admission, and subsequently implement preventative measures [6,7,8,9,10,11,12,13,14]. With the accumulation of large amounts of medical data and the development of machine learning techniques, research on machine learning prediction models for predicting delirium has been stimulated [15,16,17,18,19,20,21,22,23,24,25,26].

Machine learning techniques such as the decision tree ensemble do not need data preprocessing (neither normalization nor standardization), and can handle missing values without listwise deletion. Such models can use considerable data and capture nonlinearity and interaction. In our study on ischemic stroke patients, we confirmed that machine learning can be used to predict prognoses more accurately than using previously developed prognostic scores [27]. However, the “black box” problem, whereby recognizing processes and formulas is difficult, hinders the medical implementation of machine learning-based methods.

In a previous study, by focusing on the high prediction accuracy of machine learning technology and the disease management function of electronic clinical pathways, we developed a method for the management of patients with cerebral hemorrhage after stratifying the risk of developing aspiration pneumonia based on machine learning prediction results [28]. However, owing to the black-box nature of machine learning, the predictive contributing factors could not be well visualized, and the routine use of machine learning technology in clinical workflows was not possible.

Machine learning interpretation methods (ML interpretation methods), which have attracted research attention in recent years, can enable clinicians to interpret predictive models and provide appropriate feedback to analysts, promoting the development of data-driven medicine. Therefore, the use of ML interpretation methods for clinical data has been actively investigated [29,30,31,32]. However, because these methods have been developed recently, there are no examples of their operational incorporation into clinical workflows. Although there are previous studies that have also applied interpretation methods to models for predicting delirium [24,25,26], there are no examples of their implementation in clinical workflows.

Therefore, this study aims to build a data-driven decision support system by incorporating the predictive results of machine learning into daily clinical workflows, using a ML interpretation method. In other words, the objective of this study is to demonstrate that it is possible to develop a learning health system for delirium using an interpretation method.

2. Materials and Methods

The transparent reporting of a multivariable prediction model for individual prognosis or diagnosis was used in developing the prediction model [33].

2.1. Data Source and Transparency

We extracted clinical variables from clinical data accumulated through routine clinical practice in the hospital database of Saiseikai Kumamoto Hospital, and created an analytical database. The data, the methods used in the analysis, and the materials used to conduct the research will be made available to any researcher for the purpose of reproducing the results or replicating the procedure. Owing to the sensitive nature of the data collected for this study, only requests to utilization the dataset from qualified researchers trained in human subject confidentiality protocols will be sent by the corresponding author.

2.2. Study Design and Patients

We analyzed data from patients hospitalized at Saiseikai Kumamoto Hospital, Kumamoto City, Japan. This hospital is an emergency and critical care center in southern Japan. The study was approved by the institutional review board (approval no. 1072). Here, 61,181 consecutive patients hospitalized between December 2017 and February 2022 were studied. Of the 5792 patients excluded from this study, 2963 patients could not be traced, 357 had died within 24 h after admission, 2223 had delirium on admission, and 249 were less than 18 years old. A total of 55,389 patients were considered in this analysis. Furthermore, for focusing on cases prone to delirium, patients in the analysis were categorized into two groups (Figure 1).

2.3. Clinical Outcomes

Incident delirium was defined with a positive confusion assessment method (CAM) [34] after 24 h of hospitalization. The occurrence of postoperative delirium was defined with a positive CAM after leaving the operating room. Patients were assessed daily by nurses trained by a specialized delirium subcommittee that included a psychiatrist.

2.4. Predictors

All predictors were obtained prior to the onset of delirium. Predictive models were developed for the three populations subject to analysis, and various predictors were used for each (

Table 1. Predictor characteristics in analysis population A.

Predictors	Overall n = 55,389	No Delirium n = 50,928	Delirium n = 4461	p Value	% Missing
Predisposing risk factors
Age, year	73.0 (64.0–82.0)	72.0 (64.0–81.0)	83.0 (74.0–89.0)	<0.001	0.0
Body mass index, kg/m²	22.5 (20.0–25.1)	22.7 (20.2–25.3)	20.6 (18.2–23.2)	<0.001	1.1
Men	33,829 (61.1)	31,474 (61.8)	2355 (52.8)	<0.001	0.0
Intake of benzodiazepine medications	977 (1.8)	812 (1.7)	165 (3.8)	<0.001	4.4
Intake of opioid medications	896 (1.7)	690 (1.4)	206 (4.7)	<0.001	4.4
Intake of steroid medications	1829 (3.5)	1573 (3.2)	256 (5.9)	<0.001	4.4
Dementia	4907 (9.3)	3450 (7.1)	1457 (33.3)	<0.001	4.4
Brain tissue disorder	7028 (13.3)	5950 (12.3)	1078 (24.7)	<0.001	4.4
Heavy drinker	878 (1.7)	701 (1.4)	177 (4.0)	<0.001	4.4
History of delirium	1307 (2.5)	802 (1.7)	505 (11.5)	<0.001	4.4
Emergency hospitalization	30,043 (54.2)	26,075 (51.2)	3968 (88.9)	<0.001	0.0
Use of ambulance	19,189 (34.6)	15,948 (31.3)	3241 (72.7)	<0.001	0.0
Room at hospitalization				<0.001	0.0
Bay of general ward	16,139 (29.1)	15,570 (30.6)	569 (12.8)
Intensive care unit	12,216 (22.1)	9661 (19.0)	2555 (57.3)
Private room of general ward	27,034 (48.8)	25,697 (50.5)	1337 (30.0)
Schedule of treatment				<0.001	0.0
Catheterization	7249 (13.1)	6950 (13.6)	299 (6.7)
Endoscopic treatment	5032 (9.1)	4792 (9.4)	240 (5.4)
Preserved treatment	23,554 (42.5)	21,421 (42.1)	2133 (47.8)
Surgery	19,554 (35.3)	17,765 (34.9)	1789 (40.1)

Data are expressed as median (interquartile range), number (%).

, Tables S1 and S2). In population A, which included all patients, 14 predictors were used. In population B, which included emergency hospitalization patients, 44 predictors were used. In population C, which included elective general anesthesia surgical patients, 23 predictors were used. When developing models for all patients, only basic patient information available on admission, such as age, gender, history of delirium, and treatment plan after admission was used. For emergency hospitalization and elective general anesthesia surgical patients, subject-specific predictors such as laboratory data, vital signs, and intraoperative information were used.

2.5. Development of Prediction Models

Herein, eXtreme gradient boosting [XGBoost], a type of decision tree-based ensemble learning algorithm, was used to develop the prediction model [35]. This algorithm was selected because of its high prediction accuracy, its ability to capture non-linear relationships between outcomes and predictors, and its compatibility with the ML interpretation method used herein, as described later. The details of the machine learning programs are presented in Supplementary File S1.

2.6. Handling of Missing Data

For predictors with missing data, we used the sparsity-aware split-finding method in XGBoost, which treats the missing data as informative as they are.

2.7. Assessment of Prediction Models

The discriminating ability of each model was assessed from the area under the receiver operating characteristic (AUROC) curve as well as that under the precision–recall curve (AUPRC), in addition to the sensitivity and specificity. The AUPRC was calculated because it is an insightful performance metric for imbalanced data such as those for infrequent outcome events [36]. The Youden index method was used to determine the optimal cut-off value that maximizes sensitivity and specificity. The calibration performance of each model was assessed using the calibration slope and intercept. The overall performance was assessed using the Brier score [37]. A stratified five-fold cross-validation was performed for internal validation, and the mean and standard deviation of each metric were calculated.

2.8. Confirmation of Predictive Contributors Using the Mchine Learning Interpretation Method

We used Shapley additive explanations (SHAP), a model-agnostic method of machine learning, to visualize the relationship between predictors and delirium. SHAP is an excellent method for visualizing the contribution of predictors. Shapley values are guaranteed to be fairly distributed in cooperative game theory [38]. The SHAP algorithm is an additive feature attribution method that approximates each prediction $f\left(x\right)$ with $g\left(z^{\prime }\right)$, a linear function of the binary variables $z^{\prime }\in {\left\{0,1\right\}}^M$ and the feature attribution values ${\varphi }_i\in ℝ$, which are defined as follows:

$$g\left(z^{\prime }\right)={\varphi }_0+{\displaystyle \sum }_{i=1}^M{\varphi }_i{z}_i^{\prime },$$

(1)

where $M$ is the number of predictors.

The SHAP value ${\varphi }_i$ is defined as follows:

$${\varphi }_i={\displaystyle \sum }_{S\subseteq \mathrm{N}\backslash \left\{i\right\}}\frac{\left|S\right|\mathsf{!}\left(M-\left|S\right|-1\right)\mathsf{!}}{M\mathsf{!}}\left[f_x\left(S\cup \left\{i\right\}\right)-f_x\left(S\right)\right],$$

(2)

where $f$ is the model, $N$ is the set of all input variables, and $S$ is a subset of set $N$ excluding variable $i$. The importance of each variable visualized in the SHAP summary plots is defined as follows:

$$\mathit{FeatureImportanc}{e}_i=\frac{1}{N}{{\displaystyle \sum }}_{j=1}^N\left|{\varphi }_i^{\left(j\right)}\right|,$$

(3)

where $j$ is the subscript of the case number. The XGBoost algorithm was used to visualize Shapley values because of its high affinity to the SHAP method.

2.9. Statistical Methods

The differences in baseline characteristics and clinical data were compared using the χ² test, Fisher’s exact test, and the Mann–Whitney U test, as appropriate. Two-sided probability values less than 0.05 were considered statistically significant. All statistical analyses were conducted using the R statistical package (http://www.r-project.org/, version 4.2.0).

3. Results

3.1. Patient Characteristics

The median age (and IQR) of the 55,389 patients was 73.0 (64.0–82.0) years, and 33,829 (61.1%) of them were men. Delirium occurring 24 h or more after admission was reported among 4461 (8.1%) of the 55,389 patients in population A, and among 3968 (13.2%) of the 30,043 patients in population B. Postoperative delirium occurred in 206 (4.8%) of 4293 elective general anesthesia surgical patients. Table 1 lists the differences in predictors with and without delirium in population A; all the predictors differed significantly. Similarly, on examining the differences in population B, significant differences were observed in all predictors except laboratory data potassium and total bilirubin (Table S1). In population C, many predictors, except brain tissue disorder, heavy drinking, central venous port insertion, physical restraints, and the numerical rating scale, which is a pain rating scale, differed significantly (Table S2). The missing ratio of predictors was less than 10% for all the predictors except the Glasgow Coma Scale (Table 1, Tables S1 and S2).

3.2. Performance of Delirium Prediction Models Based on XGBoost

The prediction performance of the model based on XGBoost after cross-validation is presented in

Table 2. Predictive performance of prediction models.

Analysis Population	Algorithm	Discrimination				Calibration		Overall
		Sensitivity	Specificity	AUROC	AUPRC	Slope	Intercept	Brier Score
A, n = 55,389	XGBoost	0.838 (0.015)	0.721 (0.022)	0.852 (0.005)	0.329 (0.015)	1.013 (0.038)	0.001 (0.049)	0.062 (0.002)
B, n = 30,043		0.783 (0.027)	0.690 (0.015)	0.806 (0.006)	0.365 (0.014)	1.102 (0.017)	0.002 (0.017)	0.098 (0.001)
C, n = 4293		0.767 (0.084)	0.709 (0.040)	0.794 (0.033)	0.177 (0.052)	1.032 (0.347)	−0.047 (0.306)	0.044 (0.004)

The mean and standard deviation of each metric after stratified 5 fold cross-validation are shown.

. For the discrimination performance of the prediction model, the AUROC was 0.852 for population A, 0.806 for population B, and 0.794 for population C. Similarly, the AUPRC calculated to assess the ability to adequately identify imbalanced events was 0.329 in population A, 0.365 in population B, and 0.177 in population C. Furthermore, the calibration slope was 1.013 for population A, 1.102 for population B, and 1.032 for population C. Discrimination and calibration performances were generally good.

3.3. Interpretation of Predictors

The SHAP summary plots visualized age, body mass index (BMI), dementia, and history of delirium as key predictors in all analyzed populations (Figure 2, Figure 3 and Figure 4). Other predictors specific to each population were visualized. We identified SHAP dependence plots to visualize the relationship between the BMI and albumin levels with delirium, which were not focused on in the pre-revision risk score used at the acute care hospital in this study but made a large contribution to prediction (Figure 5 and Figure 6). Furthermore, to review the reference value of age in the pre-revised risk score, the relationship between age and delirium was similarly visualized (Figure 7). The results revealed a sharp increase in risk when BMI was near 20–25 and albumin levels were near 3.0–3.5. The risk of delirium increased as both values decreased. The risk of delirium increased rapidly in the age group of 70–75 years.

3.4. Review and Use of Analysis

An expert panel including a psychiatrist reviewed the results of this analysis and consented to add the BMI and albumin levels as new risk factors. Furthermore, they increased the risk threshold from age 60 to 70 years (Figure 8). The revised risk score can lower the false positive and false negative rates by 4.66% and 0.02%, respectively. A novel clinical pathway was introduced on 3 August 2022. The pathway includes the revised risk score and delirium-specific care content.

4. Discussion

The major findings of this study are as follows. The discriminative performance and calibration of the XGBoost-based model were satisfactory. The visualization of SHAP highlighted the effect of the BMI and albumin levels on delirium and the cut-off values of age (Figure 5, Figure 6 and Figure 7). These factors have previously received limited attention, and their values were reflected in the revision of risk scores.

Although the populations and predictors used in previous studies on the prediction of delirium using machine learning differ from those used in the present study, they exhibited higher AUROCs of 0.78–0.95 [15,16,17,18,19,20,21,22,23,24,25,26]. However, most studies were conducted on Western subjects, and the results may differ for Asians. The current study yielded a certain value because of the numerous cases involving Asians. Because Japan has an aging population, many of its hospitalized patients are senior citizens. Therefore, risk factors for delirium may not necessarily match those of other countries. Comparing the risk factors presented in previous studies with those extracted in the current study, many factors such as age, functional status, cognitive impairment, and dementia are common [4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,39,40,41]. However, reviewing the predictive contributing factors visualized in SHAP revealed that low BMI and albumin levels were risk factors for delirium (Figure 5 and Figure 6). This result may reflect a low nutritional level associated with thinness, i.e., frailty. Thus, the development and utilization of individualized predictive models that reflect the unique domestic situation may exhibit certain benefits.

In a previous study, we focused on the high prediction accuracy of machine learning and the disease management capability of electronic clinical pathways, and confirmed a deterrent effect on aspiration pneumonia by revising care items such as oral care in the electronic pathways for hemorrhagic stroke patients, using data analysis [28]. The combination of machine learning and electronic clinical pathways could be used to realize a learning health system; however, the black box nature of machine learning hinders its application. The ML interpretation method can enable experts to interpret the effect of various predictors of delirium and allow health care providers to make data-driven decisions. Figure 9 shows a schematic of a series of clinical workflows that incorporate the results of analyses based on ML interpretation methods. The analyst submits analysis results obtained using the interpretation method in an expert meeting on delirium. Specifically, the expert meeting is held based on the results, and if the results are judged to have clinical value, they are used in actual clinical practice. Note that an expert meeting including a psychiatrist was conducted to validate the results of the analysis. SHAP and other ML interpretation methods are still in the nascent stage of development, and the interpretation of visualized results should be kept in mind. These results should only be used to detect anomalies in the input data or to understand the effect of predictors. Against this background, this study was limited to revising the risk score with reference to the results of the ML interpretability method.

However, revising the risk score alone cannot accurately reflect the complex interactions among predictors and cannot improve the prediction accuracy. The false-positive rate would remain high even with a revised risk score. Therefore, predictive models based on machine learning must be implemented in the future; for example, in patient information systems for electronic medical records. Furthermore, a system should be constructed to initiate measures against delirium based on the predictive values for each case output of the models. Studies focused on the prospective implementation of machine learning-based delirium prediction have exhibited consistency with independent expert ratings; however, calibration challenges have been reported [42]. In addition, the predictive performance of machine learning predictive models has been reported to deteriorate across hospitals [43], leaving challenges for the development and implementation of generic machine learning predictive models that can be used across different hospitals. To spread the implementation of medical diagnosis support systems using machine learning prediction models, it is necessary to seriously consider whether prediction models specific to each hospital should be developed by each hospital, or whether prediction models should be developed using integrated data from different hospitals after establishing a statistical mechanism to eliminate inter-institutional bias. Furthermore, to implement predictive models using machine learning, in addition to the challenges of the predictive models, operational problems such as the timing of data acquisition should be resolved. Therefore, further research is necessary.

This study had several limitations. First, patients from a single center were selected, which resulted in potential selection bias. Second, delirium incidence was low, and data including such unbalanced events may affect model development. Third, although the prediction models were internally validated, external validation was not performed, owing to limited time. Therefore, the validity of the prediction accuracy was not assured. Finally, because this study was conducted on Japanese subjects, the results of the analysis should be assessed in other social settings and ethnic groups.

5. Conclusions

Although research on machine learning predictive models using ML interpretation methods is on the rise, studies showing the sequence of events leading up to actual implementation in clinical practice are rare. In this study, machine learning models were developed for predicting delirium for numerous Japanese cases. The models displayed high degrees of discrimination and calibration performance. A risk score was revised using the ML interpretation method, and an example of incorporating the analysis results of electronic health record data into a clinical workflow was presented.