Asthma affects 8.4% of the US population [1], resulting in over 3000 deaths, almost 500,000 hospitalizations, and over 2,000,000 emergency department (ED) visits every year [1,2]. The state-of-the-art method for reducing asthma-related hospital visits, including ED visits and hospitalizations, is to use a model to forecast which of the patients with asthma are prone to future poor outcomes. We then enroll these patients in care management, ensuring care managers call them periodically to help them schedule health and related services. As used by health care systems like University of Washington Medicine, Intermountain Healthcare, and Kaiser Permanente Northern California [3] and by health plans in 9 of 12 metropolitan communities [4], this method, if implemented properly, can cut up to 40% of patients’ future hospital visits [5-8].

Having a limited capacity, a care management program can enroll only a small portion of patients [9], and its effectiveness is upper bounded by the accuracy of the predictive model. Because they are missing certain important features, the existing models for forecasting asthma-related hospital visits among patients with asthma [3,10-22] are inaccurate, with each model missing over half of patients who will undergo future asthma-related hospital visits and mislabeling many others as making such visits. As a result, care management programs continue to be used inefficiently because they are unable to focus on the highest-risk patients. In addition, patient outcomes deteriorate, whereas health care costs increase. To address this problem, we recently considered many candidate features and developed two extreme gradient boosting (XGBoost) [23] machine learning models, one on Intermountain Healthcare data [24] and the other on Kaiser Permanente Southern California (KPSC) data [25], to forecast asthma-related hospital visits in the succeeding 12 months among patients with asthma with a higher accuracy. As is typical for machine learning approaches, these two models do not explain their forecasting results. Clinicians would know that a patient is considered to be at a high risk by the model, but the model offers no reason why this is the case. This makes it difficult for clinicians to understand and trust the model’s prediction result, determine whether the patient should be put into care management, and pinpoint interventions suitable for the patient. To address the interpretability issue of black-box models, we designed an automatic method to offer rule format explanations for any machine learning model’s forecasting results on imbalanced tabular data and to suggest customized interventions with no accuracy loss [26]. Our method worked well for explaining our Intermountain Healthcare model’s forecasting results [26], but its generalizability to other health care systems remains unknown.

The objective of this study is to evaluate the generalizability of our automatic explanation method to KPSC in forecasting asthma-related hospital visits. In the following sections, we describe our evaluation approach and results.

After receiving approval from the institutional review boards of KPSC and University of Washington Medicine, in this study, we conducted a secondary analysis of retrospective data.

We adopted the same patient cohort from our previous KPSC predictive model paper [25]: a random sample of 70% of patients with asthma who had a KPSC health plan for any period between 2015 and 2018. This sample size is the largest one permitted by KPSC for sharing its data with another non-Kaiser Permanente institution for research. KPSC has 227 clinics and 15 hospitals. It is the largest integrated health care system in Southern California, offering care to approximately 19% of residents there [27]. A patient was deemed asthmatic in a specific year if during that year, at least one asthma diagnosis code (International Classification of Diseases, Tenth Revision [ICD-10]: J45.x; International Classification of Diseases, Ninth Revision [ICD-9]: 493.0x, 493.1x, 493.8x, and 493.9x) was recorded on the patient in the encounter billing database [11,28,29]. Patient death during that year served as the exclusion criterion.

We adopted the same prediction target as our prior KPSC predictive model paper [25]. For every patient deemed to have asthma in a specific year, the indicator of any asthma-related hospital visit in the succeeding year is the outcome. An asthma-related hospital visit is a hospitalization or ED visit with asthma as its principal diagnosis (ICD-10: J45.x; ICD-9: 493.0x, 493.1x, 493.8x, and 493.9x). When training and testing our automatic explanation method and our KPSC XGBoost model, for each patient who had a KPSC health plan on a year’s last day and was also deemed asthmatic in the year, we used the patient’s data up to the year’s last day to forecast the patient’s outcome in the succeeding year.

We adopted the same administrative and clinical data set from our prior KPSC predictive model paper [25]. Obtained from KPSC’s research data warehouse, this structured data set covered our patient cohort’s visits at KPSC between 2010 and 2018.

Our KPSC model [25] uses the XGBoost classification algorithm [23] and 221 features to forecast asthma-related hospital visits in the succeeding year in patients with asthma. These features are listed in our previous KPSC predictive model paper [25], were computed on the structured attributes in our data set, and cover various characteristics such as patient demographics, medications, visits, diagnoses, vital signs, procedures, and laboratory tests. An example feature is the total number of asthma relievers that the patient filled in the previous 12 months. Every input data instance to our KPSC model aims at a (patient, index year) pair, includes these 221 features, and is used to forecast the succeeding year’s outcome of the patient. As in our prior KPSC predictive model paper [25], the top 10% of patients with asthma projected at the highest risk were used as the cutoff point for binary classification. We used the same data preprocessing approach adopted in our prior KPSC predictive model paper [25] to clean, normalize, and prepare the data.

Previously, we designed an automatic method to offer rule format explanations for the forecasting results of any machine learning model on tabular data and to suggest customized interventions with no accuracy loss. The original method [30] was designed for relatively balanced data. Recently, we extended the method to handle imbalanced data [26], where one value of the outcome variable has a much lower prevalence rate than another. This fits the case of forecasting asthma-related hospital visits in patients with asthma. At KPSC, the prevalence rate of having asthma-related hospital visits in the succeeding year was approximately 2%. In the remainder of this paper, we focus on the extended automatic explanation method.

In our experiments, we used the same parameter setting approach used in our previous automatic explanation paper [26]. Each association rule had no more than 5 items on its left-hand side. Our KPSC XGBoost model [25] used 221 features to forecast asthma-related hospital visits. We used the top 50 features that our KPSC model ranked with the largest importance values to construct association rules. Our KPSC model gained an area under the receiver operating characteristic curve (AUC) of 0.820 using all 221 features and an AUC of 0.815 using the top 50 features.

For forecasting asthma-related hospital visits, our KPSC model [25] obtained a lower AUC for KPSC data than our Intermountain Healthcare model on Intermountain Healthcare data [24]. As mentioned in our previous automatic explanation paper [26], the harder it is to forecast the outcome, the smaller the minimum commonality and confidence thresholds need to be to ensure that our automatic explanation method can provide explanations for a large percentage of the patients whom the first model correctly forecasts to take a poor outcome value. Following this guideline on KPSC data, we set the minimum commonality threshold to 0.08%, which is lower than the corresponding value of 0.2% we used on Intermountain Healthcare data [26]. We set the minimum confidence threshold to 25%, which is lower than the corresponding value of 50% used for the Intermountain Healthcare data [26]. Despite not looking large, 25% is much greater than 2%, which is the percentage of KPSC data instances associated with asthma-related hospital visits in the succeeding year, as well as our KPSC model’s positive predictive value of 11.03% [25].

To set the value of the confidence difference threshold τ, we calculated the number of association rules passing the rule pruning process versus τ. Our previous paper [26] shows that this number of rules first drops quickly as τ rises and then drops slowly when τ becomes sufficiently large. The value of τ was set at the transition point.

Remember that each data instance aims at a different (patient, index year) pair. Tables 1 and 2 present the demographic and clinical characteristics of our KPSC patient cohort during 2012-2016 and 2017, respectively. The two sets of characteristics are sufficiently similar to each other. During 2012-2016, 2.42% (18,925/782,762) of data instances were linked to asthma-related hospital visits in the succeeding year. During 2017, this fraction was 2.13% (4353/204,744). Our previous KPSC predictive model paper [25] provides a detailed comparison of the two sets of characteristics.

Taking the top 50 features that our KPSC model ranked with the largest importance values, we mined 11,628,850 association rules from the training set. Figure 1 displays the number of residual rules versus the confidence difference threshold τ. This number first drops quickly as τ rises and then drops slowly when τ becomes ≥0.15. Accordingly, the value of τ was set to 0.15, resulting in 954,493 residual rules.

An asthma clinical expert in our team labeled the values and value ranges of the top 50 features that could have a positive correlation with asthma-related hospital visits in the succeeding year. After we removed those rules involving any other value or value range, 725,632 association rules remained. Each rule provides a reason why a patient is forecasted to undergo future asthma-related hospital visits. Almost all (725,623) of these rules were actionable. Thus, our automatic explanation method’s performance numbers are almost the same regardless of whether all these rules or only the actionable rules were used. In the remainder of this section, we present only the performance numbers when only the actionable rules were used.

To allow the reader to gain a sense of the association rules the second model adopted, we present 5 example rules:

The use of many asthma relievers indicates poor asthma control. An intervention tied to the item the patient filled ≥89 asthma relievers in total in the prior 12 months is to tailor prescribed medications and to suggest the patient to maximize adherence to asthma control medications or to improve avoidance of asthma triggers.

The use of many nebulizer medications indicates a poor asthma control. An intervention tied to the item the patient had ≥25 nebulizer medication orders in the prior 12 months is to tailor prescribed medications and to suggest the patient to maximize adherence to asthma control medications or to improve the avoidance of asthma triggers.

As defined in our previous paper [24], major visits for asthma cover outpatient visits linked to a primary diagnosis of asthma, and ED visits and hospitalizations linked to an asthma diagnosis code. Outpatient visits linked to a secondary, but not a primary, diagnosis of asthma are deemed minor visits for asthma. Having many major visits for asthma indicates a poor asthma control. An intervention tied to the item the patient incurred ≥16 major visits for asthma in the prior 12 months is to adopt control strategies for the patient to avoid needing emergency care.

Having many primary or principal asthma diagnoses indicates a poor asthma control. An intervention tied to the item the patient had 8 or 9 primary or principal asthma diagnoses in the prior 12 months is to offer the patient suggestions on how to improve asthma control.

Having many no shows correlates with poor outcomes. An intervention tied to the item the patient had ≥6 no shows in the prior 12 months is to give the patient social resources to handle socioeconomic challenges to keep appointments.

In the United States, Black and African American people tend to have poorer asthma outcomes than others. Frequent ED visits indicated a poor asthma control. An intervention tied to the item the patient incurred ≥8 ED visits in the prior 12 months is to adopt control strategies for the patient to avoid needing emergency care.

Short-acting beta-2 agonists are rescue medications for the quick relief of asthma symptoms. The use of many short-acting beta-2 agonists indicates a poor asthma control. An intervention tied to the item the patient was ordered ≥28 short-acting beta-2 agonist medications in total in the prior 12 months is to tailor prescribed medications and to suggest the patient to maximize adherence to asthma control medications or to improve the avoidance of asthma triggers.

Status asthmaticus is the most severe form of asthma exacerbation. An intervention tied to the item the highest exacerbation severity of all of the asthma diagnoses recorded on the patient in the prior 12 months is status asthmaticus is to offer the patient suggestions on how to improve asthma control.

Having many visits with same-day appointments indicates a poor asthma control. An intervention tied to the item the patient incurred ≥11 and ≤17 visits with same day appointments in the prior 12 months is to improve support offered to the patient between visits to enhance medication adherence, address asthma triggers, and maximize the value of each visit.

A patient incurs a nonelective visit when the patient’s condition requires an immediate medical attention, for example, when the patient experiences severe asthma exacerbation. An intervention tied to the item the admission type of the patient’s last visit in the prior 12 months is nonelective is to adopt control strategies for the patient to avoid needing emergency care.

We evaluated our automatic explanation method on the test set. Our method explained the forecasting results for 97.9% (599/612) of the children (age <18 years) with asthma and 97.45% (1605/1647) of the adults (age ≥18 years) with asthma our KPSC model correctly forecasted to undergo asthma-related hospital visits in the succeeding year. Put together, our method explained the forecasting results for 97.57% (2204/2259) of the patients with asthma who were correctly forecasted to undergo asthma-related hospital visits in the succeeding year. For every such patient, on average, our method provided 1516.25 (SD 2161.30) explanations, each from one rule, and found 24.04 (SD 8.68) actionable items.

For the patients with asthma whom our KPSC model correctly forecasted to undergo asthma-related hospital visits in the succeeding year, Figures 2 and 3 display the patient distribution according to the number of actionable rules suiting a patient. Having a long tail, this distribution is significantly skewed toward the left. As the number of rules suiting a patient increases, the number of patients each covered by this number of rules tends to decline nonmonotonically. The biggest number of rules suiting a patient is fairly large (15,252). However, only one patient matched this number of rules.

For the patients with asthma whom our KPSC model correctly forecasted to undergo asthma-related hospital visits in the succeeding year, Figure 4 displays the patient distribution according to the number of different actionable items included in all the rules suiting a patient. The largest number of different actionable items included in all the rules suiting a patient is 42, much less than the largest number of actionable rules suiting a patient. As noted in our previous automatic explanation paper [26], 2 or more actionable items included in the rules that suit a patient often connect to the same intervention.

Our automatic explanation method provided explanations for 67.61% (2943/4353) of patients with asthma who would undergo asthma-related hospital visits in the succeeding year.

The results presented in this paper are similar to those presented in our previous automatic explanation paper [26]. For forecasting asthma-related hospital visits, our automatic explanation method exhibited an acceptable generalizability to KPSC. In particular, our method explained the forecasting results for 97.57% (2204/2259) of the patients with asthma who were correctly forecasted to undergo asthma-related hospital visits in the succeeding year. This fraction is comparable with that (89.68%) on Intermountain Healthcare data in our previous automatic explanation paper [26] and is large enough to put our automatic explanation method into daily clinical use. After further development to boost its accuracy, our KPSC model combined with our automatic explanation method could be used to guide asthma care management’s use to help enhance patient outcomes and reduce health care costs.

Our automatic explanation method provided explanations for 67.61% (2943/4353) of patients with asthma who would undergo asthma-related hospital visits in the succeeding year. This fraction is less than the 97.57% (2204/2259) success rate, at which our method explained the forecasting results for the patients with asthma our KPSC model correctly forecasted to undergo asthma-related hospital visits in the succeeding year. This is possibly due to the correlation between the association rules’ and our KPSC model’s forecasting results. Among the patients with asthma whom our KPSC model correctly forecasted to undergo asthma-related hospital visits in the succeeding year, many are easy cases for us to explain their outcomes using association rules. Among the patients with asthma who would undergo asthma-related hospital visits in the succeeding year and whose outcomes were incorrectly forecasted by our KPSC model, many are difficult cases for any model to correctly explain or forecast their outcomes.

Many rules could suit a patient. In this case, it is undesirable to list all of them simultaneously and overwhelm the user of the automatic explanation function. Instead, we should rank these rules and display the top few (eg, 3) of them by default. If desired, the user can ask an automatic explanation function to show more rules. In ranking the rules suiting a patient and the items on the left-hand side of a rule, we consider the following factors and strike a balance among them:

We are in the process of preparing a paper describing our rule-ranking method in detail.

As described in the book [34] and the survey paper [35], many other researchers have proposed miscellaneous methods for automatically offering explanations for the forecasting results of machine learning models. Such explanations are typically not in a rule format. Many such methods sacrifice a part of the forecasting accuracy and/or are designed for a particular machine learning algorithm. In addition, none of these methods can automatically suggest customized interventions. In comparison, our automatic explanation method supplies rule format explanations for any machine learning model’s forecasting results on tabular data and suggests customized interventions with no accuracy loss. Rule format explanations are easier to comprehend and can suggest customized interventions more directly than other forms of explanations.

To the best of our knowledge, we were the first to use association rules to automatically offer rule format explanations for any machine learning model’s forecasting results on tabular data and to suggest customized interventions with no accuracy loss [30]. Our original method [30] was designed for relatively balanced data and was initially tested in the case of forecasting type 2 diabetes diagnoses. Subsequently, Alaa et al [36,37] applied the original method to multiple medical prediction tasks. So far, no researcher outside of our group has applied our extended automatic explanation method [26], which can handle imbalanced data, to any prediction task. Rudin et al [38] and Ribeiro et al [39] used rules to automatically offer explanations for the forecasting results of any machine learning model. These rules are not association rules and are unknown before the prediction time; hence, they cannot be used to automatically suggest customized interventions at the prediction time. In comparison, the association rules used in our automatic explanation method are mined before the prediction time and used to automatically suggest customized interventions at the prediction time.

This study has three limitations, all of which can be fine areas for future work:

In its first generalizability assessment, our automatic explanation method for imbalanced tabular data exhibited a decent generalizability to KPSC for forecasting asthma-related hospital visits. After further development to boost its accuracy, our KPSC model combined with our automatic explanation method could be used to guide asthma care management’s use to help enhance patient outcomes and reduce health care costs.