This study employed a retrospective electronic health record (EHR) chart review of over 75 factors known to be relevant for PrI risk in individuals with SCI and routinely recorded in the EHR. We also perform tissue health assessments for a selected sub-group. Regulatory approval for the study was obtained from the local institutional review board. By applying a data-centric approach, we can leverage the power of the data resource provided by VINCI and the detailed personal characteristic database of tissue health to provide the weighted, adaptive, personalized SCIPUD+ Resource for primary and secondary PrI prevention.

The integrated SCIPUD+ Resource is being assembled from 2 databases: (1) one using data extracted from the VINCI EHR by informatics and text mining, and (2) another using tissue health data. PrI risk factor data collected at multiple retrospective time points include modifiable and nonmodifiable factors identified in cross-sectional and observational studies. Multiscale data extraction includes numerical, categorical and text data mining. A Spinal Cord Injury Pressure Ulcer and Deep Tissue Injury Ontology (SCIPUDO), is being developed to ensure robust and extensive information extraction from the free text clinical note. We will also carry out a cross-sectional study of tissue health profiles in a representative cohort of 60 individuals with SCI.

The Multi-Modality, Multi-Resource Information Integration Environment for Multi-Center Physiological and Clinical Research Studies (Physio-MIMI) cloud-based multi-modal data storage and access platform [44] creates a common Web-based user interface for data queries. It also enables the development of compatible analytical tools and easier sharing of complex data from multiple domains to support collaborative clinical and translational research using diverse data types. Another tool, Ontology-driven Web-based Research Data Capture (OnWARD) provides robust flexibility of input data storage in a relational database for detailed analysis. It can be quickly deployed and customized for any clinical study. OnWARD has eased the data entry burden in multiple clinical trials [45].

Structural modeling of factors from multiple domains and their co-impact on developing PrI will be used to provide weighted systemic insight into initial and recurrent PrI risk in people with SCI. A comprehensive model will be used to relate the primary outcome of interest (ie, PrI development) with covariates including environmental, social, clinical, personal, and tissue health profiles and possible interactions among some of these covariates.

The SCIPUD+ Resource is being developed using a detailed chart review of VINCI data employing International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes for paraplegia and tetraplegia with a secondary filter using an SCI-specific stop code. The search timeframe is the preconversion date (September 2010 to September 2015) because the International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) codes do not currently provide accurate delineation of SCI factors. The initial query returned approximately 36,000 different individuals and 120,000 encounters during the search timeframe across the VHA nationally. It was clarified that some individuals coded for SCI have a primary diagnosis of multiple sclerosis (MS) or amyotrophic lateral sclerosis (ALS). Therefore, we revised the code to develop a secondary filter to exclude individuals with MS and ALS since risk factors vary considerably in these neurodegenerative diseases compared to SCI. This secondary query revealed a study cohort of over 20,000 individuals with SCI, equivalent to about 8% of the total United States population. Within this cohort, a detailed review found that it includes more than 109,000 encounters. Furthermore, we have learned that each encounter encompasses an episode of care and may include different appointments stemming from the same visit, or an extended period of hospitalization. Thus, we have estimated the cohort includes about 500,000 different events and over 40 million data points.

Our research strategy builds on our existing methodologies to create the SCIPUD+ Resource to enable personalized care planning for PrI prevention based on the individual’s holistic characteristics [41]. Analysis of multiple PrI risk factors requires a robust and scalable informatics approach to cope with challenges in volume and complexity. Clinical and demographic data is collected using systems with a variety of sampling rates and formats. Even when checklists and coding are required, data may be missing or only found in the free form note. During preliminary work, we found that ICD-9-CM codes markedly under-reported the number of PrI treated. In a population of 399 eligible patients, only 93 (23.3%) were coded for PrI. We have developed a pathway for construction of disease-specific ontologies for data extraction using Natural Language Processing (NLP) for complex, specialized clinical notes. We will create the dedicated domain ontology SCIPUDO by reusing terminology from existing systems ranging from anatomy (Systematized Nomenclature of Medicine-Clinical Trials), disease classification (ICD-9-CMand ICD-10-CM), medication (website RxNorm for clinical trial drug standardized nomenclature), and NINDS CDE. Due to Physio-MIMI’s highly adaptable system architecture with domain ontology as a plug-and-play component, the proposed SCIPUD+ Resource can be developed by reusing much of the existing open-source tools that we have already developed. Figure 2 shows 2 hypothetical examples. In the first scenario, the clinical profile and tissue health response are the most critical domains. Potentially modifiable factors in these domains include spasticity and applied loads. Thus, the SCIPUD+ care plan would prioritize spasticity management and equipment provided. In the second scenario, the critical domains are personal and clinical factors. In addition to the potentially modifiable factors in the clinical domain, potentially modifiable factors in the personal domain include smoking and body mass index.

Based on our prior data, we defined the expected PrI incidence as 30% and a clinically significant difference as reducing the incidence by 50%. The basic PrI status extracted from the EHR is PrI or not PrI, leading to a dichotomous outcome. However, the severity of PrI differs. We will use text mining to further classify wound status as severe PrI (stage 3 or 4), minor PrI (stage 1 or 2), deep tissue injury, absent or unclassified, leading to a polytomous outcome. A first-line analysis model for dichotomous outcome uses logistic regression, while the first-line analysis model for a polytomous outcome uses multinomial logistic regression. Considering all variables and their possible interactions would lead to approximately 3082 covariates to be studied in each model. In practice, it is reasonable to expect that only a small portion of these covariates, possibly as few as 25 would be enough to predict PrI outcomes. Only clinically meaningful interactions need to be considered at the start of our modeling. To achieve a reasonably rich SCIPUD+ database that allows for a balanced cohort selection of the personal, environmental, social, and clinical factors and for an extensive study of the impact of these factors, we can and will oversample. Therefore, data will be extracted from a retrospective chart review of 5000 individuals selected from the study cohort of over 36,000 individuals with SCI using a stratified sampling scheme. We will retain 500 representative cases for further validation and testing. This sampling will provide more than 1418 (ie, 5000 minus 3082 minus 500) degrees of freedom, which is more than enough to determine the top 25 predictors. It will also validate and test these predictors with at least 80% power under a standard significance level of 0.05, assuming an average difference of PrI incidence of at least 0.15, and a moderately balanced number of cases [46-48]. The validation and testing of these top 25 predictors will be based on standard tests and bootstrap procedures.

Raw data is extracted from VINCI using a stratified ICD-9-CM code search, de-identified and stored as a comma-separated values (CSV) file. Data formats include categorical, numerical and free-form text in clinical notes. Data to be collected includes factors identified as being possibly related to PrI development or healing either in cross-sectional or other observational studies. Annual evaluations for the complete study cohort will be included in the SCIPUD+ database, to determine changes over time. Any patient admission will also be reviewed, together with weekly in-patient and discharge encounters. We are particularly interested in the possible differences in PrI risk based on the level and extent of spinal injury and motor and sensory impairment. Thus, we will examine quadriplegia motor-complete (QMC), quadriplegia motor-incomplete (QMI), paraplegia motor-complete (PMC), and paraplegia motor-incomplete (PMI).

We are developing SCIPUDO as the knowledge resource for processing specialized terms related to SCI, PrI, and deep tissue injuries. Parsing and analyzing clinical narratives present a unique set of challenges that distinguish it from the broader biomedical NLP approaches. There has been extensive work in creating clinical NLP systems focused on information extraction from free text in specific disease domains, such as cancer [49] and tuberculosis [50]. However, there is no community gold standard for SCI or PrI named entities to date. Thus, we will build our gold standard using manual annotations created by clinical team members who will review a random sample of records from over 20,000 clinical notes for this cohort extracted from VINCI. One to 3 clinicians will review each record. The SCIPUDO will enable: (1) term disambiguation (ie, between commonly used synonyms and acronyms of a term such as quadriplegia and tetraplegia), (2) term normalization (ie, syntactic variations of a term, such as singular or plural and acronyms will be normalized using classes and customized rules such as pressure ulcer, PU, or PrU), and (3) subsumption reasoning using class hierarchy to allow terms to be classified according to their broader semantic type.

Validated extracted data will then be collated using our established standard data collection forms and uploaded to the Physio-MIMI based integrated PrI risk assessment SCIPUD+ Resource. The Physio-MIMI backbone will provide extensible, scalable, and high-performance data management for storing and accessing large volumes of data rapidly. Reliable data storage through automated data replication and data integrity verification will ensure consistent data availability and effective disaster recovery with off-site data backup. Data quality assurance and metadata version control will be managed using a combination of GitHub, JSON, and the open source NSR data management environment [51].

Input data for the SCIPUD+ database will be provided by synthesizing available EHR clinical data from VINCI, using a protocol based on our preliminary work. VINCI provides EHR data storage for all health care encounters within the VHA and is updated daily. A preliminary query on June 10, 2014, found that between 2009-2014 there were 16,076 individuals seen by the VHA with an ICD-9-CM code of 344.00 (ie, quadriplegia) and 24,052 individuals with an ICD-9-CM code of 344.1 (ie, paraplegia). Of these, 6420 (16.00%) in both groups were also coded for a PrI (ICD-9-CM code 707.00). Within the local area, there were 1021 encounters with individuals with quadriplegia and 1443 with paraplegia. The reported rate of PrI incidence was 14.00%. Extraction of clinical details will entail text mining of the free text clinical notes. SCIPUDO will enable robust text mining for data extraction from free form notes in addition to using ICD-9-CM codes to retrieve data of interest. A visual query interface will be adapted from OnWARD to allow all clinicians to directly query the SCIPUD+ Resource via a set of readily usable visual widgets that will be populated with the SCIPUDO classes to allow clinicians to construct queries, specific to the patient flexibly. All patient data will be stored in a firewall protected secure environment with role-based access control and audit trail logging.

We will develop the SCIPUD+ user interface which will provide a single point of Web-based access to well-annotated and de-identified data generated from multiple domains. Modifiable and nonmodifiable factors will be considered (Figure 1). To develop the SCIPUD+ environmental, social and clinical PrI risk structural model we will consider PrI status as the response variable. We will employ general logistic and multinomial logistic models with linear mixed effects (transformed if necessary) and interaction terms to fit the data. Tree-based models such as classification and regression tree (CART) and Random Forest will also be used to examine the relationship of the factors to the PrI status. Model and variable selection will be implemented to define the SCIPUD+ environmental, social, and clinical model. Final models will be validated using cross-validation. Both models, especially the tree-based models are useful to rank-order factors to identify specific critical variables for an individual, with a focus on modifiable factors (Figure 2).

General logistic and multinomial logistic models with linear (mixed) effects and their possible interaction terms will then be fit to the data or their natural groups, and the significance of PrI development will be assessed. Natural groupings will be obtained using a cluster analysis of 5000 medical records and 60 detailed tissue health profiles to examine the association of these natural grouping with PrI frequency. The representativeness of the tissue health group will be compared with the larger sample drawn from the larger cohort extracted from VINCI EHR. The integrated model will be developed in the same way as the SCIPUD+ environmental, social and clinical structural model. Using statistical software R and Splus, tree-based models such as CART and Random Forest, will also be used to examine the relationship of all factors with PrI status. Model and variable selection based on both logistic and tree-based models will be implemented to define the integrated SCIPUD+ model. Final models will be validated using 10-fold cross-validation and some hold out cases using predictive measures. Both models, especially the tree-based models are useful to rank-order factors for identification of specific critical variables for an individual. We shall pay particular attention to modifiable factors. The comprehensive model proposed will allow us to borrow the degrees of freedom from all data points to develop the fully integrated SCIPUD+ Resource. We will determine structural models based on both data sources using standard statistical models, and directly using large-p small-n modern techniques for all factors. Special interest models, such as those focused on modifiable factors, will also be developed. The choice of essential features will depend on the optimization criteria used by a model fitting or learning algorithm. For example, the Random Forest provides 2 criteria for ranking the important features, also known as the variable of importance. One is based on the contribution to Model Accuracy and the other to Gini impurity by each included variable [52-54]. We will use the domain knowledge to guide our final choice of the model for different medical purposes if the final models differ significantly based on various criteria. We may also use XGboost as needed to derive and validate the best predictive model [55]. We expect that at most 25 top-ranked factors will be enough for modeling PrI risk as either a dichotomous or polytomous outcome. This will allow development of the SCIPUD+ care planning algorithm.