About 8.4% of people in the United States have asthma [1], which causes over 3000 deaths, around 500,000 hospitalizations, and over 2 million emergency department (ED) visits each year [1,2]. To improve patient outcomes and cut the number of asthma-related hospital encounters including ED visits and hospitalizations, predictive models are widely used to prospectively pinpoint high-risk patients with asthma for preventive care via care management. This is the case with health care systems such as the University of Washington Medicine, Kaiser Permanente Northern California [3], and Intermountain Healthcare, and with other health plans in 9 of 12 metropolitan communities [4]. Once a patient is identified as high risk and placed into a care management program, a care manager will call the patient periodically to assess asthma control, adjust asthma medications, and make appointments for needed care or testing. Successful care management can help patients with asthma obtain better outcomes, thereby avoiding up to 40% of their future hospital encounters [5-8].

A care management program has a limited service capacity and usually enrolls ≤3% of patients [9] with a given condition, which places a premium on enrolling at-risk patients. Therefore, the accuracy of the adopted predictive model (or lack thereof) puts an upper bound on the effectiveness of the program. Previously, several researchers have developed several models for projecting asthma-related hospital encounters in patients with asthma [3,10-22]. Each of these models would consider only a few features, miss more than half of patients who will have future asthma-related hospital encounters, and incorrectly project future asthma-related hospital encounters for many other patients with asthma [23]. These errors lead to suboptimal patient outcomes, including hospital encounters and unnecessary health care costs because of unneeded care management program enrollment. When building machine learning models on nonmedical data, people often follow the modeling guideline of checking extensive candidate features to boost model accuracy [24-27]. Adopting this modeling guideline to the medical domain, we recently constructed a machine learning model on Intermountain Healthcare data to project asthma-related hospital encounters in the following 12-month period in patients with asthma [23]. Compared with previous models, our model boosts the area under the receiver operating characteristic curve (AUC) by at least 0.049 to 0.859. Although this is encouraging, it remains unknown whether our modeling guideline is generalizable to other health care systems.

This study aims to assess the generalizability of our modeling guideline to Kaiser Permanente Southern California (KPSC). Similar to our Intermountain Healthcare model [23], our KPSC model uses administrative and clinical data to project asthma-related hospital encounters (ED visits and hospitalizations) in patients with asthma. The categorical dependent variable has 2 possible values—whether the patient with asthma will have asthma-related hospital encounters in the following 12-month period or not. This study describes the construction and evaluation of our KPSC model.

In this study, we performed a secondary analysis of computerized administrative and clinical data. This study was approved by the institutional review boards of the University of Washington Medicine and KPSC.

As shown in Figure 1, our patient cohort was based on patients with asthma who were enrolled in a KPSC health plan for any duration between 2015 and 2018. Owing to internal regulatory processes, the patient cohort was restricted to a random sample of 70.00% (397,858/568,369) of eligible patients. This sample size is the maximum that KPSC allows for sharing its data with an institution outside of Kaiser Permanente for research. As the largest integrated health care system in Southern California with 227 clinics and 15 hospitals, KPSC offers care to approximately 19% of Southern California residents [28]. A patient was deemed to have asthma in a particular year if the patient had one or more diagnosis codes of asthma (International Classification of Diseases [ICD], Tenth Revision [ICD-10]: J45.x; ICD, Ninth Revision [ICD-9]: 493.0x, 493.1x, 493.8x, 493.9x) recorded in the encounter billing database in that year [11,29,30]. The exclusion criterion was that the patient died during that year. If a patient had no diagnosis code of asthma in any subsequent year, the patient was deemed to have no asthma in that subsequent year.

For each patient identified as having asthma in a particular year, the outcome was whether the patient had any asthma-related hospital encounter in the following year. An asthma-related hospital encounter is an ED visit or hospitalization with asthma as the principal diagnosis (ICD-10: J45.x; ICD-9: 493.0x, 493.1x, 493.8x, 493.9x). For every patient with asthma, the patient’s data up to the end of every calendar year were used to project the patient’s outcome in the following year as long as the patient was deemed to have asthma in the previous year and was also enrolled in a KPSC health plan at the end of the previous year.

For the patients in our patient cohort, we used their entire electronically available patient history at KPSC. At KPSC, various kinds of information on its patients has been recorded in the electronic medical record system since 2010. In addition, we had electronic records of the patients’ diagnosis codes starting from 1981, regardless of whether they were stored in the electronic medical record system. From the research data warehouse at KPSC, we retrieved an administrative and clinical data set, including information regarding our patient cohort’s encounters and medication dispensing at KPSC from 2010 to 2018 and diagnosis codes at KPSC from 1981 to 2018. Owing to regulatory and privacy concerns, the data set is not publicly available.

We examined 2 types of candidate features—basic and extended. A basic feature and its corresponding extended features differ only in the year of the data used for feature computation. We considered 307 basic candidate features listed in Multimedia Appendix 1 [31]. Covering a wide range of characteristics, these basic candidate features were computed from the structured attributes in our data set. In Multimedia Appendix 1, unless the word different shows up, every mention of the number of a given type of item such as medications counts multiplicity. As defined in our previous paper [23], major visits for asthma include ED visits and hospitalizations with an asthma diagnosis code and outpatient visits with a primary diagnosis of asthma. Outpatient visits with a secondary but no primary diagnosis of asthma is regarded as minor visits for asthma.

Every input data instance to the model targets a unique (patient, index year) pair and is employed to forecast the patient’s outcome in the following year. For the (patient, index year) pair, the patient’s primary care provider (PCP), age, and home address were computed as of the end of the index year. The basic candidate features of history of bronchiolitis, the number of years since the first asthma-coded encounter in the data set, premature birth, family history of asthma, and the number of years since the first encounter for chronic obstructive pulmonary disease in the data set were computed using the data from 1981 to the index year. All of the allergy features and the features derived from the problem list were computed using the data from 2010 to the index year. One basic candidate feature was computed using the data in the index and preindex years: the proportion of patients who had asthma-related hospital encounters in the index year out of all of the patients of the patient’s PCP with asthma in the preindex year. The other 277 basic candidate features were computed using the data in the index year.

In addition to the basic candidate features, we also checked extended candidate features. Our Intermountain Healthcare model [23] was built using the extreme gradient boosting (XGBoost) machine learning classification algorithm [32]. As detailed in Hastie et al [33], XGBoost automatically computes the importance value of every feature as the fractional contribution of the feature to the model. Previously, we showed that ignoring those features with importance values <0.01 led to a little drop in model accuracy [23]. Using the basic candidate features and the model construction method described below, we built an initial XGBoost model on KPSC data. As a patient’s demographic features rarely change over time, no extended candidate feature was formed for any of the basic demographic features. For each basic candidate feature that was nondemographic, was computed on the data in the index year, and had an importance value 0.01 in the initial XGBoost model, we computed 2 related extended candidate features, one using the data in the preindex year and another using the data in the year that was 2 years before the index year. The only difference between the extended candidate features and the basic feature is the year of the data used for feature computation. For instance, for the basic candidate feature number of ED visits in 2016, the 2 related extended candidate features are the number of ED visits in 2015 and the number of ED visits in 2014. In brief, we formed extended candidate features for only those suitable and important basic candidate features. Our intuition is that among all possible ones that could be formed, these extended candidate features are most promising with regard to additional predictive power. For the other basic candidate features with lower importance values, those extended candidate features that could possibly be formed for them tend to have little extra predictive power and can be ignored. Given the finite data instances available for model training, this feature extending approach avoids a large rise in the number of candidate features, which may cause sample size issues. We considered all of the basic and extended candidate features when building our final predictive model.

Every data instance targets a unique (patient, index year) pair. Multimedia Appendix 1 displays the clinical and demographic characteristics of our patient cohort during the time periods of 2012 to 2016 and 2017. The set of characteristics during 2012 to 2016 is similar to that during 2017. During 2012 to 2016 and 2017, 2.42% (18,925/782,762) and 2.13% (4353/204,744) of data instances were associated with asthma-related hospital encounters in the following year, respectively.

Table 2 shows for each clinical or demographic characteristic, the statistical test results on whether the data instances linking to future asthma-related hospital encounters and those linking to no future asthma-related hospital encounter had the same distribution. These 2 sets of data instances had the same distribution when the P value is ≥.05, and distinct distributions when the P value is <.05. In Table 2, all of the P values <.05 are marked in italics.

Before building our final model, the importance values of the basic candidate features were computed once on our initial XGBoost model. This led to us examining 30 extended candidate features in addition to the 307 basic candidate features. With these 337 basic and extended candidates features as inputs, our automatic model selection method [37] picked the XGBoost classification algorithm [32]. As an ensemble of decision trees, XGBoost can handle missing feature values naturally. Our final predictive model was built using XGBoost, and the 221 features shown in descending order of importance value in Multimedia Appendix 1. The other features had no additional predictive power and were automatically dropped by XGBoost.

On the KPSC test set, our final model achieved an AUC of 0.820 (95% CI 0.813-0.826). Figure 2 displays the ROC curve of our final model. Table 3 displays the performance measures of our final model when various top percentages of patients having the largest predicted risk were adopted as the cutoff point for performing binary classification. When this percentage was at 10.00% (20,474/204,744), our final model achieved an accuracy of 90.08% (184,435/204,744; 95% CI 89.95-90.21), a sensitivity of 51.90% (2259/4353; 95% CI 50.44-53.42), a specificity of 90.91% (182,176/200,391; 95% CI 90.78-91.03), a PPV of 11.03% (2259/20,474; 95% CI 10.59-11.46), and an NPV of 98.86% (182,176/184,270; 95% CI 98.81-98.91). Table 4 gives the corresponding error matrix of our final model.

When we excluded the extended candidate features and considered only the basic candidate features, the AUC of our model dropped to 0.809. Several basic candidate features, such as the number of years since the first asthma-coded encounter in the data set, needed over one year of past data to calculate. When we further excluded these multiyear candidate features and considered only those basic candidate features calculated on 1 year of past data, the model’s AUC dropped to 0.807.

Without precluding any feature from being considered, the model trained on data from both children (aged <18 years) with asthma and adults (aged ≥18 years) with asthma gained an AUC of 0.815 in children with asthma and an AUC of 0.817 in adults with asthma. In comparison, the model trained only on data from children with asthma gained an AUC of 0.811 in children with asthma. The model trained only on data from adults with asthma gained an AUC of 0.818 in adults with asthma.

If we adopted only the top 25 features shown in Multimedia Appendix 1 with an importance value ≥0.01 and removed the other 312 features, the model’s AUC dropped from 0.820 to 0.800 (95% CI 0.793-0.808). When the top 10.00% (20,474/204,744) of patients having the largest predicted risk were adopted as the cutoff point for doing binary classification, the model’s accuracy dropped from 90.08% (184,435/204,744) to 89.96% (184,185/204,744; 95% CI 89.83-90.08), sensitivity dropped from 51.90% (2259/4353) to 49.02% (2134/4353; 95% CI 47.71-50.55), specificity dropped from 90.91% (182,176/200,391) to 90.85% (182,051/200,391; 95% CI 90.72-90.97), PPV dropped from 11.03% (2259/20,474) to 10.42% (2134/20,474; 95% CI 10.03-10.86), and NPV dropped from 98.86% (182,176/184,270) to 98.80% (182,051/184,270; 95% CI 98.75-98.85).

When applying our simplified Intermountain Healthcare model trained on the Intermountain Healthcare data set [23] to the KPSC test set without retraining the model on the KPSC training set, the model gained an AUC of 0.751 (95% CI 0.742-0.759). When the top 10.00% (20,474/204,744) of patients having the largest predicted risk were adopted as the cutoff point for doing binary classification, the model achieved an accuracy of 89.64% (183,531/204,744; 95% CI 89.51-89.77), a sensitivity of 41.51% (1807/4353; 95% CI 40.14-42.97), a specificity of 90.68% (181,724/200,391; 95% CI 90.55-90.81), a PPV of 8.83% (1807/20,474; 95% CI 8.44-9.23), and an NPV of 98.62% (181,724/184,270; 95% CI 98.57-98.67).

After using the KPSC training set to retrain our simplified Intermountain Healthcare model [23], the model gained on the KPSC test set an AUC of 0.779 (95% CI 0.772-0.787). When the top 10.00% (20,474/204,744) of patients having the largest predicted risk were adopted as the cutoff point for doing binary classification, the model achieved an accuracy of 89.85% (183,953/204,744; 95% CI 89.71-89.97), a sensitivity of 46.36% (2018/4353; 95% CI 44.89-47.84), a specificity of 90.79% (181,935/200,391; 95% CI 90.65-90.91), a PPV of 9.86% (2018/20,474; 95% CI 9.45-10.25), and an NPV of 98.73% (181,935/184,270; 95% CI 98.68-98.78).

We used KPSC data to develop a model to forecast asthma-related hospital encounters in the following 12-month period in patients with asthma. Table 5 shows that, compared with the models formerly built by others [3,10-22], our final KPSC model gained a higher AUC, that is, our modeling guideline of checking extensive candidate features to boost model accuracy exhibited acceptable generalizability to KPSC. After further enhancement to automatically explain its predictions [40,41] and to raise its accuracy, our model could be used to direct asthma care management to help improve patient outcomes and reduce health care costs.

Asthma affects adults and children differently. Our final model gained a lower AUC in children than in adults. Additional work is required to understand the difference and to boost the prediction accuracy in children.

We examined 337 basic and extended candidate features. Approximately 65.6% (221/337) of these were used in our final model. Many of the unused features were correlated with the outcome variable but provided no additional predictive power on the KPSC data set beyond those used in our final model.

In Multimedia Appendix 1, the 8 most important features and several others within the top 25 features reflect the loss of asthma control. This loss of asthma control could be because of the severity of the patient’s asthma. It could also relate to management practices, treatment nonadherence, or socioeconomic factors for which we had no data.

When using our simplified Intermountain Healthcare model [23] without retraining it on the KPSC training set, the model achieved an AUC of 0.751 on the KPSC test set. Despite being 0.069 lower than our final KPSC model’s AUC, this AUC is higher than the AUCs of many previous models for predicting hospitalization and ED visits in patients with asthma (Table 5). Therefore, we regard our simplified Intermountain Healthcare model to have acceptable generalizability to KPSC.

Multiple researchers have built models to forecast ED visits and hospitalizations in patients with asthma [3,10-23]. Table 5 compares our final KPSC model with those models, which encompass all pertinent models covered in the systematic review of Loymans et al [18]. With the exception of our Intermountain Healthcare model [23], every model formerly built by others [3,10-22] gained a lower AUC than our final KPSC model. Instead of being for all patients with asthma, the model by Miller et al [15] targets adults with difficult-to-treat or severe asthma, 8.5% of whom had future asthma-related hospital encounters. The model by Loymans et al [10] predicts asthma exacerbations with a prevalence rate of 13%. These 2 prevalence rates of the undesirable outcome are much higher than that in our KPSC data set. In addition, the target patient population and the prediction target of these 2 models are not comparable with those in our KPSC model. Except for these 2 models, each of the other models formerly built by others had an AUC ≤0.79, which is at least 0.030 lower than that of our KPSC model.

Compared with other models, the model by Yurk et al [17] gained a larger PPV and sensitivity mainly because of the use of a distinct prediction target: hospital encounters or one or more days lost because of missed work or reduced activities for asthma. This prediction target was easier to predict, as it occurred in 54% of the patients with asthma. If the model by Yurk et al [17] were used to predict asthma-related hospital encounters that occurred with approximately 2% of the patients with asthma, we would expect the model to gain a lower sensitivity and PPV.

Excluding the model by Yurk et al [17], all of the other models formerly built by others had a sensitivity ≤49%, which is smaller than what our final KPSC model gained: 51.90% (2259/4353). Sensitivity provides, among all patients with asthma who will have future asthma-related hospital encounters, the proportion of patients that the model pinpoints. As the population of patients with asthma is large, for every 1% increase in the identified proportion of patients with asthma who would have future asthma-related hospital encounters, effective care management could help improve patient outcomes, thereby avoiding up to 7200 more ED visits and 1970 more hospitalizations in the United States annually [1,5-8].

The PPV depends substantially on the prevalence rate of undesirable outcomes [42]. In our KPSC test data set, 2.13% (4353/204,744) of patients with asthma had future asthma-related hospital encounters. When the top 10.00% (20,474/204,744) of patients having the largest predicted risk were adopted as the cutoff point for performing binary classification, the maximum possible PPV that a perfect model could obtain is 21.26% (4353/20,474). Our final KPSC model gained a PPV of 11.03% (2259/20,474), which is 51.90% (2259/4353) of the maximum possible PPV. In comparison, in our Intermountain Healthcare test data set, 4.22% of patients with asthma had future asthma-related hospital encounters [23]. Our Intermountain Healthcare model gained a PPV of 22.65% (436/1925) [23], which is 53.7% (436/812) of the maximum possible PPV that a perfect model could obtain. On a data set in which 6.5% of patients with asthma had future asthma-related hospital encounters, the model by Schatz et al [20] gained a PPV of 22.0%. On a data set in which 6.9% of patients with asthma had future asthma-related hospital encounters, the model by Lieu et al [19] gained a PPV of 18.5%. Except for these PPVs and the PPV of the model by Yurk et al [17], none of the previously reported PPVs was more than 5.6%.

Despite being built using the same modeling guideline, our final KPSC model gained a lower AUC than our Intermountain Healthcare model [23]. This is largely because the percentage of data instances in the test set linking to future asthma-related hospital encounters differs greatly at Intermountain Healthcare and at KPSC: 4.22% (812/19,256) versus 2.13% (4353/204,744), respectively. The rarer the undesirable outcome, the harder it is to accurately predict it.

The top features with an importance value ≥0.01 in our final KPSC model are similar to those in our Intermountain Healthcare model [23]. In both our final KPSC and our Intermountain Healthcare models, many top features involve asthma medications and previous ED visits. When building our Intermountain Healthcare model, we did not consider several basic candidate features. They turned out to be top features in our final KPSC model and impacted the importance values and ranks of the other top features there.

When building our Intermountain Healthcare model, we did not incorporate any extended candidate features. Several such features appeared as top features in our final KPSC model. Their inclusion boosted the model accuracy on our KPSC data set. It is possible that including extended candidate features could also boost the model accuracy on our Intermountain Healthcare data set. This could be explored in future work.

Schatz et al [20] showed that in 2 Southern California cities, 6.5% of patients with asthma at KPSC had asthma-related hospital encounters in 2000. In comparison, 2.08% (4353/208,959) of patients with asthma at KPSC had asthma-related hospital encounters in 2018. This suggests that compared with 2 decades ago, KPSC manages patients with asthma better now.

Although more accurate than those formerly built by others, our final KPSC model still gained a somewhat low PPV of 11.03% (2259/20,474). However, our model could be clinically useful:

Our final KPSC model used 221 features. Cutting this number could facilitate the clinical deployment of the model. In this regard, if one could bear a small drop in prediction accuracy, one could adopt the top features having an importance value of, for example, 0.01 or more and remove the others. The importance value of a feature changes across health care systems. Ideally, before deciding which features to keep, one should first compute the importance values of the features on a data set from the intended health care system.

Most of the attributes that we used to compute the features adopted in our final KPSC model, particularly the top features, are routinely collected by electronic medical record systems. For future work, to make it easy for other health care systems to reuse our final KPSC model, we can resort to the Observational Medical Outcomes Partnership (OMOP) common data model [43]. This data model and its linked standardized terminologies [44] standardize administrative and clinical attributes from at least 10 large US health care systems [45,46]. We can extend this data model to include the attributes that are used in our final KPSC model but missed by the original data model. We rewrite our feature construction and model building code based on the extended OMOP common data model and post our code and the related data schema on a public website. After converting its data into our extended OMOP common data model format based on this data schema, a health care system can rerun our code on its data to obtain a simplified version of our final KPSC model tailored to its data. Hopefully, most of the predictive power of our final KPSC model can be retained similar to what this study showed for our Intermountain Healthcare model.

It is difficult to interpret an XGBoost model employing many features globally, as is the case with many other involved machine learning models. As an interesting topic for future work, we plan to use our previously proposed method [40,41] to automatically explain our final KPSC model’s predictions for each patient with asthma.

Our final KPSC model was an XGBoost model [32]. When classifying 2 unbalanced classes, XGBoost employs a hyperparameter scale_pos_weight to balance their weights [47]. To maximize the AUC of our KPSC model, our automatic model selection method [37] changed scale_pos_weight from its default value to balance the 2 classes of having future asthma-related hospital encounters or not [48]. As a side effect, this shrank the model’s projected probabilities of having future asthma-related hospital encounters to a large extent and made them differ greatly from the actual probabilities [48]. This does not affect the identification of the top few percent of patients with asthma who have the largest projected risk to receive care management or other preventive interventions. We could keep scale_pos_weight at its default value of 1 and not balance the 2 classes. This would avoid the side effect but drop the model’s AUC from 0.820 to 0.817 (95% CI 0.810-0.824).

This study has 3 limitations, all of which provide interesting areas for future work:

In its first generalizability assessment, our modeling guideline of examining extensive candidate features to help boost model accuracy exhibited acceptable generalizability to KPSC. Compared with the models formerly built by others, our KPSC model for projecting asthma-related hospital encounters in patients with asthma gained a higher AUC. At present, predictive models are widely used as a core component of a decision support tool to prospectively pinpoint high-risk patients with asthma for preventive care via care management. After further enhancement, our KPSC model could be used to replace the existing predictive models in the decision support tool for better directing asthma care management to help improve patient outcomes and reduce health care costs.