With the rapid development of the Semantic Web, machines are able to understand the contextual meaning of data, including in the eld of automated semantics-driven statistical reasoning. This paper introduces a semantics-driven automated approach for solving population health problems with descriptive statistical models. A fusion of semantic and machine learning techniques enables our semantically-targeted analytics framework to automatically discover informative subpopulations that have subpopulation-speci c risk factors signi cantly associated with health conditions such as hypertension and type II diabetes. Based on our health analysis ontology and knowledge graphs, the semanticallytargeted analysis automated architecture allows analysts to rapidly and dynamically conduct studies for di erent health outcomes, risk factors, cohorts, and analysis methods; it also lets the full analysis pipeline be modularly speci ed in a reusable domain-speci c way through the usage of knowledge graph cartridges, which are application-speci c fragments of the underlying knowledge graph. We evaluate the semanticallytargeted analysis framework for risk analysis using the National Health and Nutrition Examination Survey and conclude that this framework can be readily extended to solve many di erent learning and statistical tasks, and to exploit datasets from various domains in the future.
Population health strives to improve the health outcomes of subject groups
through the analysis of enormous health-related datasets collected from
members of these groups [
mons License Attribution 4.0 International (CC BY 4.0). statistics will be required to monitor and improve population-wide health situations. To understand the relationships between population health determinants and outcomes, observational studies are performed on large patient databases.
These databases include electronic health records and ongoing
populationwide surveys, such as the National Health and Nutrition Examination Survey
(NHANES, [
In this work, we present a framework, semantically targeted analytics (STA),
for automatically generating population health statistical analyses. In order to
overcome limitations on study scope, we develop a semantic representation for
knowledge from key domains: survey design and analysis, health, and data
analytics. Integration of each of these domains via a consistent standard [
When subject cohorts are no longer manually chosen, there is no guarantee
that a linear statistical model will be su cient to explain associations found
in population health datasets. Thus, we utilize the supervised cadre model
(SCM, [
In STA, semantics encodes, captures, and isolates the domain knowledge needed to model study de nitions, statistical techniques, and data. A key component is the knowledge graph cartridge (hereafter cartridges): an applicationspeci c subgraph of an underlying KG. Cartridges, further described in section 3.2, are a way to express special-purpose, application-speci c sub-graphs, to augment the graph for analysis. They are implemented as RDF KGs and enable an automated \plug and play" architecture. Further, cartridges are used either as input when analysts choose to load them to perform a novel risk study, or as output when the study nding is automatically written into them. Our cartridges are sub-graphs that contribute to a larger analysis graph. Additionally, our output cartridges de ne the results in a way that is consistent with the input cartridges and contribute to the modularity of STA.
To model and represent the components of our cartridges, we built a Health Analytics Ontology (HAO2). HAO models the domain knowledge, analytics
while and enumerating the possible classes and relations of these entities. [
knowledge, and other analytics pipeline components necessary for population health analysis.
The main contributions of this paper are a semantic representation of population health analysis work ows and results as knowledge graph cartridges, the integration of this representation with precision machine learning techniques for the discovery subpopulation-speci c risk factors, and the demonstration of how the STA framework enables rigorous investigation of population health problems. Via cartridges, our STA framework can analyze, interpret, and report studies performed on a wide variety of chronic health conditions and potential risk factors. In Section 5, we present and examine the discoveries found by applying STA to the task of subpopulation-speci c identi cation of risk factors associated with prediabetes and increased total cholesterol levels. Our framework successfully identi es risk factors that are not picked up by standard population-level risk analysis. 1.1
Our primary inspiration for the KG cartridge is the Oracle database systems
notion of a data cartridge [
HAO is inspired by several existing analytics-focused ontologies, including
the Data Science Ontology (DSO) associated with the semantic ow graph [
With a similar goal as Automated Machine Learning (AutoML, [
Algorithm 1 illustrates how the STA framework uses semantic structures realized as cartridges to drive precision health subpopulation discovery and risk factor identi cation. In STAGE I, the STA framework queries its input cartridges to infer requirements for data preparation. This might entail ltering records for subjects that satisfy study inclusion criteria, log-transforming right-skewed variables, or constructing new variables based on supplied de nitions. In STAGE II, an array of SCMs is trained on the prepared cohort using di erent hyperparameter con gurations. A nal model is determined by supplied model selection metrics, e.g., the Bayesian Information Criterion (BIC). In STAGE III, surveyweighted generalized linear models (GLMs) are trained on each discovered subpopulation. The regression coe cients and log-odds ratios estimated by these GLMs quantify the association between the supplied risk factor and response variable. After STAGE II and STAGE III, model ndings and subpopulation characteristics are written to output cartridges for future reference.
In Algorithm 1, we perform precision risk analysis for a single risk factor. In
practice, we repeat this process for many di erent categories of risk factors,
yielding a precision environment-wide association study (EWAS, [
We present the STA framework for addressing population health problems. By varying models, variables, and the underlying datasets, we can adapt this work ow for other tasks. Classi cation and multiple regression models are used to identify potential risk factors { covariates that are strongly associated with the response variable in the study cohort, after controlling for known confounders.
NHANES is constructed with a multistage complex survey design (CSD, [
Minimal necessary analysis components from the HAO are stored in modular serializations called cartridges. A cartridge is a subgraph containing applicationand analysis-speci c entities. Cartridges can be edited to include additional elements, thereby enabling the exibility to address a range of problems with STAGE I: DATA PREPARATION
Select all subjects satisfying cohort-cartridge's inclusion criteria and store as cohort Query parameters-cartridge and risk-factor-cartridge for preprocessing techniques and apply to cohort Query response-cartridge and risk-factor-cartridge for necessary control-variables Query risk-factor-cartridge for risk-factor Query response-cartridge for response-variable Query parameters-cartridge for model-selection metric Calculate population-level summary statistics of cohort and write to subpopulation-cartridge STAGE II: SUBPOPULATION DISCOVERY for every hyperparameter con guration in parameters-cartridge do Train SCM on cohort using control-variables and risk-factor to predict response-variable
Calculate SCM's metric value end
Identify SCM with optimal metric value and write its optimal hyperparameters and parameters to model-cartridge
Take optimal SCM and write to model-cartridge, serialized as pickle STAGE III: RISK MODELING for every subpopulation discovered by SCM do
Select members of cohort belonging to subpopulation Calculate summary statistics of subpopulation and write to subpopulation-cartridge Train survey-weighted GLM on subpopulation using control-variables and risk-factor to predict response-variable Extract risk-factor's p-value, regression coe cient, and regression coe cient standard error from GLM and write to results-cartridge end Algorithm 1: STA SCM analysis for subpopulation-speci c or precision risk analysis of a single risk factor. Implemented variants include examining many potential risk factors in succession via an EWAS, as well as the addition of STAGE IV: REPORT GENERATION, in which the output cartridges are used to automatically create a report describing the ndings. Reports use text, tables in the style of Table 3, and gures in the style of Fig. 2. minimal modi cation. In Section 5, we demonstrate this versatility. In STA, cartridges are loaded and modi ed as the user constructs their risk analysis study. In the Health Analysis Ontology (HAO), we support modeling of processes, components, models, variables and factors involved in a health analysis pipeline such as the one described in Algorithm 1. The HAO reuses classes and properties from existing ontologies, listed in Table 2, but we also found it necessary to introduce new terminology.
We represent the ontology using OWL and introduce property associations between classes using owl:Restrictions. Overall, HAO provides a vocabulary necessary to model the reusable components of an analysis (sio:Analysis) implemented by an analysis work ow (hao:AnalysisWork ow) that we store in cartridges (hao:Cartridge). Cartridges serve as containers that encode information about speci c portions of a work ow. For example, a response cartridge (hao:ResponseCartridge) contributes to a high-level overview of a model with entities (modeled via sio:hasAttribute) such as the analysis question, response variable, type of model, etc. The HAO schema allows for the representation of cartridges as named knowledge graphs in the TriG format3.
The HAO ontology only imports the SemanticScience Integrated Ontology (SIO), as we reuse several classes from SIO and utilize their object properties
to de ne associations between classes. For other terms that we reuse from large
ontologies such as the National Cancer Institute Thesaurus (NCIT), the
Statistical Methods Ontology (STATO) and the Ontology of Biological and Clinical
Statistics (OBCS), we apply the Minimum Information to Reference an
External Ontology Term (MIREOT, [
Cartridges can be grouped into two categories, input and output, with further subdivisions given in Table 1. Fig. 1 gives a high-level summary of the cartridge framework. In practice, cartridges are implemented as named graph collections (in TriG format) encapsulating instances of ontology classes that, when grouped together, represent di erent modules of an analysis work ow. Further, cartridges are constructed using terms from ontologies listed in Table 2. Domain-speci c choices (e.g., choice of confounders or cohort inclusion criteria) about cartridge contents are adapted from published studies and linked with provenance. In the case that outdated or inaccurate knowledge is retired, this provenance shows what cartridges need to be updated.
Currently, input cartridges must be manually de ned by domain specialists, but output cartridges are generated automatically after analysis. Minimal modication is needed to allow an input cartridge to be applied to a di erent analytics question. Cartridges can be edited to allow for the exible tailoring of a health analysis pipeline to discover new subpopulations (stato:0000203 - cohort), identify new outcomes or test di erent response variables (hao:TargetVariable). For example, creating a new analysis of hypertension based on a type 2 diabetes analysis requires only a simple edit of the response cartridge; the other input cartridges remain the same. We maintain analysis related concepts in HAO, and for cartridges such as the subpopulation cartridge requiring domain-speci c terminology we directly reference terms from ontologies in the eld, within the cartridge. Additionally, as shown in Fig. 1, cartridges contain links to other cartridges that were used to generate it, to allow for easy traversal of all the components of a work ow. 4
Our method for precision risk is the supervised cadre model [
en.html
Cartridge category Cartridge type Contents Input Response hAenaaltlyhsicsoncodnitcieopnts and background domain axioms necessary to model a given Cohort Isntucdluys,iownhiccrhitecraina buesecdhotosedneotenr-mthien-e yif oargaivdeanptseudbfjreocmtmeaxyistbineginsctluuddieeds in the user's Risk factor How categories of semantically-similar risk factors should be modeled
Parameters cRounlesgutroactioomnsplfeotrecchhoosseennmanoadleylsis work ow and potential hyperparameter Output Model tTrhaeinhinygp,earpnadrathmeertuerlessubsyedwthoicthraiitnisa ampopdlieeld, tthoenpeawraombseetrevraetsiotinmsates learned during
Summary statistics characterizing discovered subpopulations, Subpopulation including within-subpopulation variable means and rates Results aQnudanthtie craestpioonnsoef vsuabripaobpleuluastiinogn-rsepgerceissciodnisccooeverceidenatss,sosctiaantdioanrsdbeerrtworese,nanthdepr-ivsaklufaecstor
Table 1. Types of cartridges used in STA framework Ontology Pre x Purpose Health Analysis Ontology hao Inform analysis design, summarize analysis results for comparison, and generate reports
Children's Health Exposure Analysis Resource chear Represent the inclusion of environmental exposures in health research The Statistical Methods Ontology stato Represent concepts and properties related to statistical methods and analysis
The PROV Ontology prov Model provenance information for di erent applications and domains Ontology of Biological and Clinical Statistics obcs Represent additional biostatistics terms not in OBI DC Terms dct Specify all metadata terms maintained by the Dublin Core Metadata Initiative Simple Knowledge Organization System skos De ne the new terms in the HAO Table 2. Ontologies currently used in STA. The usage of these ontologies are described in sections 3.1 and 3.2 subpopulation-speci c and precision interchangeably. The SCM is applied during STAGE II of Algorithm 1. Subpopulations, which we call cadres, are subsets of the population de ned with respect to a cadre-assignment rule learned by the SCM. Subjects in the same cadre have the same association with a given risk factor. In STA, the chosen parameter and response cartridges set up the appropriate SCM and describe how to tune its hyperparameters. Optimal model parameters and hyperparameters are written to a model cartridge, which can be applied to novel subject records to determine their cadre.
We outline SCM for multivariate regression and binary classi cation. When trained on a set of subject records fxng RP and response values fyng, the SCM divides the observations into a set of M cadres. Each cadre m is characterized by a center cm 2 RP and a linear regression function em parameterized by weights wm 2 RP and a bias wm0 2 R. New observations x have (for multivariate regression) an aggregate regression score (e.g., a subject's expected total cholesterol level) or (for binary classi cation) an aggregate risk score (e.g., the logit of their probability of having prediabetes) given by f (x) = PM m=1 gm(x)em(x), where gm(x) is the probability x belongs to cadre m, and em(x) is the regression or risk score for x were it known to belong to cadre m. These have the form gm(x) = Pme0 e jjxjjxcmcjmj2d0 jj2d and em(x) = (wm)T x + wm0:
Here, jjzjjd = Pp jdpj(zp)2 1=2 is a seminorm parameterized by d 2 RP and
> 0 is a hyperparameter. SCM parameters are obtained by applying
stochastic gradient descent to a survey-weighted loss function based on mean squared
error or logistic loss, along with elastic net [
We present two risk analyses to identify subpopulation-speci c environmental
exposure factors associated with total cholesterol (TC) and prediabetic-or-worse
glycohemoglobin levels (prediabetes). Elevated levels of serum lipids such as
TC are recognized as risk factors for cardiovascular disease, and associations
between TC and environmental exposure levels were identi ed previously [
We chose a set of input cartridges shown in Table 3 for TC using control
variables from prior studies [
With our input cartridges, we run Algorithm 1 for every potential risk factor. Each risk factor is included in a single SCM that discovers subpopulations in the data. In STAGE II of Algorithm 1, each discovered subpopulation has its summary statistics written to a subpopulation cartridge to be stored in the KG. Characteristics of subpopulations with signi cant positive associations are visualized in Fig. 2A. In STAGE III of Algorithm 1, each subpopulation has a survey-weighted GLM trained on it, and the risk factor's regression coe cient and p-value are extracted. Due to the large number of hypothesis tests, false discovery correction is applied to these p-values before assessing signi cance at a threshold speci ed in the study's parameters cartridge (here, = 0:02).
Cartridge type Contents
TC is a continuous response variable; subjects' age, Body Mass Index (BMI), Response Poverty Income Ratio (PIR), smoking habits, drinking habits, gender, marital status, and education level should be controlled for Cohort All available NHANES subjects Risk factor 201 environmental exposure risk factors divided into 17 categories
Standardize risk factor measurements; train models with M = 1; 2 and 3 cadres Parameters and choose best one using BIC for model selection; signi cance threshold of = 0:02 for GLM hypothesis tests Table 3. Chosen input cartridges for TC risk study. The prediabetes risk study uses the same cohort, risk factor, and parameters cartridge, with a di erent response cartridge. We have presented a semantically-targeted analytics framework via which risk factors speci c to a subpopulation may be discovered in datasets. With the supervised cadre machine learning method, we simultaneously discover subpopulations and identify their signi cant risk factors. To support this we built a novel Health Analysis Ontology that captures analytics and health domain knowledge.
HAO and other ontologies provide structure for de ning cartridges that are used for modular analysis pipelines. We leverage this semantic modeling to dynamically construct and execute a risk model and interpret results. Using STA, the system provides explainable insights for future population health studies in a scienti cally rigorous and reproducible way.
In STA, statistical ndings and parameters are encoded in results cartridges and written back to the KG, enabling retrieval for further study. Cartridges provide semantic extensions that enable a KG system to apply inference to solve domain-speci c analytics problems. By publishing the results cartridges, studies become reproducible and explainable with provenance. Researchers with new analysis methods can readily compare results with prior studies, using the same work ow on the same problems. They can adapt existing peer-reviewed studies to new diseases by editing cartridge in published work ows.
We report here on semantically-targeted analytics applied to population health studies that rapidly enables new ndings from the ongoing NHANES database. Using new cartridges, STA can readily be adapted to other types of statistical analysis on other data sources such as electronic health care records.
Created with support by IBM Research AI through the AI Horizons Network.