The importance of SNOMED CT concept specificity in healthcare analytics

Abstract

Background

Healthcare data frequently lack the specificity level needed to achieve clinical and operational objectives such as optimising bed management. Pneumonia is a disease of importance as it accounts for more bed days than any other lung disease and has a varied aetiology. The condition has a range of SNOMED CT concepts with different levels of specificity.

Objective

This study aimed to quantify the importance of the specificity of an SNOMED CT concept, against well-established predictors, for forecasting length of stay for pneumonia patients.

Method

A retrospective data analysis was conducted of pneumonia admissions to a tertiary hospital between 2011 and 2021. For inclusion, the primary diagnosis was a subtype of bacterial or viral pneumonia, as identified by SNOMED CT concepts. Three linear mixed models were constructed. Model One included known predictors of length of stay. Model Two included the predictors in Model One and SNOMED CT concepts of lower specificity. Model Three included the Model Two predictors and the concepts with higher specificity. Model performances were compared.

Results

Sex, ethnicity, deprivation rank and Charlson Comorbidity Index scores (age-adjusted) were meaningful predictors of length of stay in all models. Inclusion of lower specificity SNOMED CT concepts did not significantly improve performance (ΔR² = 0.41%, p = .058). SNOMED CT concepts with higher specificity explained more variance than each of the individual predictors (ΔR² = 4.31%, p < .001).

Conclusion

SNOMED CT concepts with higher specificity explained more variance in length of stay than a range of well-studied predictors.

Implications

Accurate and specific clinical documentation using SNOMED CT can improve predictive modelling and the generation of actionable insights. Resources should be dedicated to optimising and assuring clinical documentation quality at the point of recording.

Keywords

(MeSH)projections and predictions length of stay SNOMED CT specificity statistical data analysis clinical coding health information management

Introduction

Across the healthcare economy, masses of data are generated and stored within medical records. To be useful, this information must be accessible, complete and feature high levels of accuracy, uniformity and specificity (Alonso et al., 2020a; Davoudi et al., 2015). Data specificity is an important characteristic of high-quality information and refers to the level of detail observable in data (Cirillo et al., 2021). Low data specificity is common in healthcare datasets (Alonso et al., 2020b; Roberts et al., 2016; Roberts and Cheung, 2017), and it impedes the delivery of personalised patient care, clinical and operational improvement, population health and research (Cirillo et al., 2021; Davoudi et al., 2015; Department of Health and Social Care, 2018). When data are nonspecific, distinguishing and potentially insightful attributes are lost. For example, if a patient had a diagnosis of viral tonsillitis, but only a record of tonsillitis was made in the medical record, the potentially significant information about aetiology is omitted. After all, not all cases of tonsillitis have a viral origin (Georgalas et al., 2014). Where aetiology is of importance for cohort identification, analytics and research, it would need to be determined via other means, such as using a combination of microbiological test results and other clinical notes. This process can be time-consuming, resource-intensive and error-prone.

Clinical terminologies, such as SNOMED CT (SCT), are a valuable tool for creating, managing and using meaningful and computer-processable data (Lee et al., 2014). As a result, SCT is used worldwide and is mandated for use in the United Kingdom (UK) and the United States of America (USA) (National Information Board, 2014; Ostropolets et al., 2020; SNOMED International, 2020). SCT is built around an extensive and internationally recognised collection of meaningful clinical phrases representing clinical ideas and entities (Ostropolets et al., 2020). These ideas are called concepts and cover the entire clinical domain, including diagnoses, procedures, observations, organisms, events, situations and medications (SNOMED International, 2022c). Effective use of SCT supports clinicians in efficiently capturing meaningful data in health records, including problem lists, reports, assessments and clinic letters (Roberts et al., 2022).

Every SCT concept is unique and has a logical definition outlining its characteristics and relationship to other concepts (SNOMED International, 2022c). Each concept is linked via webs of cross-hierarchical and parent–child relationships within a polyhierarchical structure (SNOMED International, 2022c). Content is more generalised higher up the hierarchies, whereas content lower down the hierarchies is more specific and expressive. As such, this structure facilitates the creation of high-quality data flows at any level of specificity desired. An example of this is a clinician recording a diagnosis of hypertension in a problem list. The clinician could record the SCT term “hypertension” and offer no further modification to their statement. The SCT concept 38341003 |Hypertensive disorder, systemic arterial (disorder)| would be recorded and stored for use (SNOMED International, 2022b). The clinician could use the SCT hierarchies to identify relevant concepts with higher specificity, such as neonatal hypertension (206596003 |Neonatal hypertension (disorder)|), or diastolic hypertension (48146000 |Diastolic hypertension (disorder)|) (SNOMED International, 2022c). SCT hierarchies can be configured in SCT-enabled systems to support “intelligent” messaging to clinicians, flagging where more specific concepts may be relevant or where there are conflicts with the information recorded elsewhere in the medical record. SCT concepts can also be represented with associated synonyms. Using the example above, they could enter one of the SCT synonyms for hypertension, such as “high blood pressure”, “hypertensive disorder”, or “systemic arterial hypertension” (SNOMED International, 2022b). All of which would map to the aforementioned concept (38341003 |Hypertensive disorder, systemic arterial (disorder)|).

SCT implementation and use is not widely standardised across organisations and sectors, even where national mandates exist (Lee et al., 2013; NHS Digital, 2020). The scale and scope of SCT integration within medical records depend on the systems used as well as the available resources and expertise (Lee et al., 2013; SNOMED International, 2022a). Some providers utilise legacy interface terminologies and cross-map terms to SCT concepts. Others use SCT directly by either utilising terminology servers (McBride et al., 2012) or integrating the SCT release files directly within their record systems (NHS Digital, 2022b). Use of terminology servers is advantageous as it utilises a single source of expertise for terminology management and also allows for use of multiple SCT extensions and other useful coding systems (McBride et al., 2012).

Inpatient length of stay (LOS) is an important measure of hospital performance and demand (Bahrmann et al., 2019; Lisk et al., 2019). Factors that affect LOS include patient demographics, socioeconomic status and medical history (Bahrmann et al., 2019; Cabre et al., 2004; Coevoet et al., 2013; Mathijssen et al., 2016; Yadav et al., 2019). Reliable predictions of LOS can enable hospitals to allocate their resources better, tailor care to the individual and set reasonable patient expectations (Baniasadi et al., 2019; Cabre et al., 2004).

Pneumonia represents a significant burden for health systems, accounting for more admissions and bed days than any other lung disease (Millett et al., 2015). This problem is expected to grow with an ageing population and increasing medical complexity (Chalmers et al., 2017; Millett et al., 2015). As the aetiology of pneumonia varies, a heterogeneous subset of SCT concepts can be used to define and record occurrence (SNOMED International, 2022b). Data on infectious respiratory diseases frequently lacks details on aetiology, and there is a need to address critical omissions (Roberts et al., 2016; Roberts and Cheung, 2017).

This study aimed to determine the relative importance of the specificity of pneumonia SCT concepts against well-known predictors of LOS. LOS was selected as a proxy for core hospital analytics, as most secondary and tertiary care settings have strategic objectives in monitoring and reducing LOS.

Method

Research design

Data were collected on pneumonia admissions to a tertiary hospital organisation in the UK between 2011 and 2021. The study organisation spanned five main hospital sites and provided care for 145,000 inpatients and day cases in 2020/21 alone (Guy’s and St Thomas’ NHS Foundation Trust, 2022).

Use of SNOMED CT

Throughout the study period, the organisation utilised a locally-derived interface terminology in the electronic problem list (∼10,000 terms). Clinicians used the terminology to record diagnoses and findings throughout an admission. The interface terms were directly cross-mapped to pre-coordinated content from the international edition of SCT (SNOMED International, 2022b). Clinicians labelled the primary diagnosis for a given admission next to a local term. When a diagnosis was recorded, the local term and its associated SCT concept were stored together, enabling the wider analytical benefits of SCT to be explored. In most cases, the local terms included SCT fully specified names but also included some synonyms. Clinicians could also add free text if the local interface terminology list did not contain the desired diagnosis. This free text could be analysed so that additions to the searchable list of clinically-relevant terms could be added and cross-mapped to SCT concepts.

The scope of diagnosis terms locally was limited compared to the range of concepts within SCT’s clinical findings hierarchy, potentially creating an overdependence on free text. The use of free text creates difficulties in correct mapping and encoding at source. This requires significant resources to manage and potentially leads to data quality issues, negatively affecting downstream activities, including reporting and analytics. Clinicians did not have access to the SCT hierarchies within the electronic health record (EHR) itself. Instead, where desired, they utilised the International Terminology browser to support hierarchy traversing (SNOMED International (2022b). It is likely that many clinicians did not take this extra step, resulting in the potential loss of specificity had the hierarchies been available in a single system. Full SCT integration is currently underway as part of a large implementation of a new EHR to maximise the use of SCT features.

Data collection

Using the SCT query language and the SHRIMP tool (SCT version: UK 19/01/2022) (CSIRO, 2022), two terminologists identified admissions with a primary diagnosis of subtypes of either bacterial or viral pneumonia (SCT: 53084003 |Bacterial pneumonia (disorder)| or 75570004 |Viral pneumonia (disorder)|). Any concept that was not a direct subtype of these concepts was excluded from the study to ensure a comparative specificity analysis was conducted (SNOMED International, 2022b). It should be noted that the majority of primary diagnoses related to pneumonia were recorded as 233604007 |Pneumonia (disorder)| at the study site. These were excluded as aetiology was unclear using the terms alone, and specificity discovery (where applicable) would have taken significant resources.

The terminologists labelled each admission with two applicable SCT concepts. The first concept was the specific SCT concept captured in the record, which included the causative agent in the term. For analysis, this represented a concept of higher specificity. Based on the parent-child relationships for the recorded subtype, a second concept was assigned as either bacterial or viral pneumonia (SNOMED International, 2022b). This concept represented a concept of lower specificity. Variables previously found to be significant predictors of LOS were also collected: gender (Lequertier et al., 2021; Mathijssen et al., 2016; Ono et al., 2010), Charlson Comorbidity Index Scores (age-adjusted) (CCIS (age-adjusted)) (Bahrmann et al., 2019), ethnicity (Bruce and Smith, 2020) and Index of Multiple Deprivation (IMD) rank (Department for Levelling Up and Ministry of Housing, 2020). These variables acted as a baseline for LOS prediction without including any pneumonia-related SCT concepts. The CCIS (age-adjusted) incorporates a patient's age which has itself been shown to be significant in LOS modelling (Charlson et al., 1994; Mathijssen et al., 2016). Charlson Comorbidity Index Scores and age are highly correlated (Charlson et al., 1994), so to avoid undesirable interactions, age was not included as a standalone variable.

Three linear mixed models were constructed. Model One used only the baseline predictors of LOS. Model Two used the predictors in Model One with the addition of the lower specificity concepts. Model Three used the predictors in Model Two with the higher specificity concepts also included. Model fitting and parametric bootstrap confidence intervals (10,000 samples) were obtained using the R package “lme4” (Bates et al., 2015). Treating patient identifiers (IDs) as a random effect, which is resampled during bootstrapping, removed the effects of pseudoreplication. LOS was log-transformed to reduce skewness, and the CCIS (age-adjusted) were log-transformed using a log(x+1) transform to reduce the leverage of extreme data points. The predictors (and their use as covariates, fixed factors and random factors) were summarised. R² values were defined based on the “marginal” and “conditional” R² values of Nakagawa and Schielzeth (Nakagawa and Schielzeth, 2013), including all random effects except for the patient ID. Approximate normality and homoscedasticity of residuals and random effects were checked.

Preliminary time stability analysis

Model stability over time is paramount to reaching accurate conclusions, so preliminary time analysis was performed to quantify and limit negative effects. The median admission date was ascertained (4th March 2017) from the initial sample of 1353 admissions. The data were split into equal halves around the median and grouped as PreMedian or PostMedian. Model estimates were analysed for stability across the two halves of the dataset, using a likelihood ratio test to compare the model with only time added versus the model in which all the interactions with time were also added. Using the whole initial dataset (PreMedian and PostMedian samples combined), the interactions with time were statistically significant (χ²(8) = 18.37, p = .019), accompanied by considerable instability in the predictions of the effects for each respective specificity. This instability in the whole dataset is shown in Figure S1(a), online supplement. The head and tail of each arrow show the predictions for the effect of each specificity on length of stay in the PreMedian and PostMedian time periods, respectively. For example, LOS prediction for Haemophilus influenzae pneumonia (HIP) admissions is 27% shorter than the average LOS using the PreMedian sample but 43% shorter using the PostMedian sample. The same analysis was performed using the PostMedian sample only, with a median admission date of 30th June 2019. This showed that the interaction of the predictors with time was not statistically significant (χ²(8) = 2.92, p = .94). The predictions for the effects of the specificities were suitably stable and, therefore, much more robust for drawing meaningful conclusions (Figure S1(b), online supplement). As such, this is the dataset that was used for further analysis.

Ethical considerations

All local governance and project approval policies were followed. The NHS opt-out service was used to exclude patients who did not want their data to be used for research and planning (NHS Digital, 2022a).

Results

The composition of the final sample (677) is broken down in Table 1. Almost a quarter of the sample (24%) were under five years old, with a second peak in the older age groups. Most patients (93%) presented only once, though a few (1%) had three to five admissions for pneumonia within the study period. The median LOS was 142 hours. The grouped categories for the assigned SCT concepts are outlined in Table 2.

Table 1.

Sample composition.

Predictor		Value	Count	% of total
Admissions per patient		1	582	93.3%
		2	35	5.6%
		3	4	0.6%
		4	2	0.3%
		5	1	0.2%
Sex		Male	371	54.8%
Sex		Female	306	45.2%
Ethnicity		White British	269	39.7%
		Not Stated	98	14.5%
		White: Any other white background	61	9.0%
		Black or Black British: African	37	5.5%
		Black or Black British: Caribbean	29	4.3%
		White: English	28	4.1%
		Other: Latin American	23	3.4%
		Black: Mixed Black	20	3.0%
		White: Irish	13	1.9%
		Asian or Asian British: Bangladeshi	12	1.8%
		Asian: Mixed	10	1.5%
		Others (each)	<10	<1.5%
Age on admission		age < 5	163	24.1%
		age < 20	228	33.7%
		20 ≤ age < 40	65	9.6%
		40 ≤ age < 60	110	16.2%
		60 ≤ age < 80	197	29.1%
		80 ≤ age < 100	75	11.1%
		100 ≤ age	2	0.3%
	Mean	SD	Min	Max
IMD rank	12135.3	7002.1	1038	32527
Age on admission	42.6	31.4	0	101
CCIS (age-adjusted)	4.1	4.1	0	22

Key: CCIS – Charlson Comorbidity Index Score; IMD Rank – Index of Multiple Deprivation Rank Note: Where one participant had multiple admissions, the mean was taken of the age and CCIS (age-adjusted) values recorded across all admissions for the patient.

Table 2.

Pneumonia SCT concepts sampled.

Lower specificity			Higher specificity
Acronym	SNOMED CT concept*	Sample size	Acronym	SNOMED CT concept*	Sample size
VP	75570004 \|Viral pneumonia (disorder)\|	311
			AP	41207000 \|Adenoviral pneumonia (disorder)\|	26
			HMPP	445096001 \|Pneumonia caused by Human metapneumovirus (disorder)\|	59
			PP	64917006 \|Parainfluenza virus pneumonia (disorder)\|	70
			RSVP	195881003 \|Pneumonia caused by respiratory syncytial virus (disorder)\|	88
			CP	882784691000119100 \|Pneumonia caused by severe acute respiratory syndrome coronavirus 2 (disorder)\|	68
					311
BP	53084003 \|Bacterial pneumonia (disorder)\|	366
			ECP	51530003 \|Pneumonia caused by Escherichia coli (disorder)\|	11
			HIP	70036007 \|Haemophilus influenzae pneumonia (disorder)\|	30
			KPP	64479007 \|Pneumonia caused by Klebsiella pneumoniae (disorder)\|	38
			MPP	46970008 \|Pneumonia caused by Mycoplasma pneumoniae (disorder)\|	45
			PSP	41381004 \|Pneumonia caused by Pseudomonas (disorder)\|	95
			STAP	441658007 \|Pneumonia caused by Staphylococcus aureus (disorder)\|	37
			STEP	34020007 \|Pneumonia caused by Streptococcus (disorder)\|	110
		677			366

Note: SCT Version 31/01/2022 (SNOMED International, 2022b).

Figure 1 illustrates the distributions of the LOS for the SCT concepts with higher specificity. The geometric means for higher specificity concepts (diamonds) show meaningful deviations from those of lower specificity (dashed lines). For the bacterial types of pneumonia, Haemophilus influenzae pneumonia (HIP) and Escherichia coli pneumonia (ECP) were associated with the largest deviations in LOS. The most notable changes for the viral pneumonia subtypes were associated with Adenoviral pneumonia (AP). The LOS for all subtypes of bacterial pneumonia differed from their grouped average. Less variation in LOS was found for subtypes of viral pneumonia.

Figure 1.

Boxplots for LOS for subtypes of bacterial and viral pneumonia. Key: Diamonds: Geometric means for higher specificity pneumonia concepts; Dashed Line: Geometric means for lower specificity pneumonia concepts; AP – 41207000 |Adenoviral pneumonia (disorder)|; CP – 882784691000119100 |Pneumonia caused by severe acute respiratory syndrome coronavirus 2 (disorder)|; ECP – 51530003 |Pneumonia caused by Escherichia coli (disorder)|; HIP – 70036007 |Haemophilus influenzae pneumonia (disorder)|; HMPP – 445096001 |Pneumonia caused by Human metapneumovirus (disorder)|; KPP – 64479007 |Pneumonia caused by Klebsiella pneumoniae (disorder)|; MPP – 46970008 |Pneumonia caused by Mycoplasma pneumoniae (disorder)|; PP – 64917006 |Parainfluenza virus pneumonia (disorder)|; PSP – 41381004 |Pneumonia caused by Pseudomonas (disorder)|; RSVP – 195881003 |Pneumonia caused by a respiratory syncytial virus (disorder)|; STAP – 441658007 |Pneumonia caused by Staphylococcus aureus (disorder)|; STEP – 34020007 |Pneumonia caused by Streptococcus (disorder)| (SNOMED International, 2022b).

Figure 2 illustrates the impact of the additional predictors in each model. The dots represent the amount that the expected LOS changed up or down when SCT concepts were added iteratively to a model. For a case with a diagnosis of pneumonia caused by Pseudomonas (PSP), the LOS prediction was revised upwards by 42%, compared to that which would arise by using only the other variables. In contrast, if admission was to treat a patient with Haemophilus influenzae pneumonia (HIP), the prediction was reduced by 45%.

Figure 2.

Changes in predicted LOS for each variable. Key: *** = p < .001; CCIS – Charlson Comorbidity Index Score (age-adjusted); IMD Rank – Index of Multiple Deprivation Rank; AP – 41207000 |Adenoviral pneumonia (disorder)|; CP – 882784691000119100 |Pneumonia caused by severe acute respiratory syndrome coronavirus 2 (disorder)|; ECP – 51530003 |Pneumonia caused by Escherichia coli (disorder)|; HIP – 70036007 |Haemophilus influenzae pneumonia (disorder)|; HMPP – 445096001 |Pneumonia caused by Human metapneumovirus (disorder)|; KPP – 64479007 |Pneumonia caused by Klebsiella pneumoniae (disorder)|; MPP – 46970008 |Pneumonia caused by Mycoplasma pneumoniae (disorder)|; PP – 64917006 |Parainfluenza virus pneumonia (disorder)|; PSP – 41381004 |Pneumonia caused by Pseudomonas (disorder)|; RSVP – 195881003 |Pneumonia caused by a respiratory syncytial virus (disorder)|; STAP – 441658007 |Pneumonia caused by Staphylococcus aureus (disorder)|; STEP – 34020007 |Pneumonia caused by Streptococcus (disorder)|(SNOMED International, 2022b). Note: Estimated/predicted effects of the variables added to each subsequent model and their 95% Cis. These values represent the impact on our expectation for a given length of stay, compared with the case where we did not use that information. The later models are nested within the earlier ones. For example, Model Three also contains all those predictors from Models One and Two as controls.

The outputs of the three models and the underlying estimates are presented in Table 3. Gender, ethnicity, CCIS (age-adjusted) and IMD rank score explained 8.25% of the LOS variance in the sample (Model One). All variables were statistically significant, with different effect sizes (Table 3). The lower specificity data items (75570004 |Viral pneumonia (disorder)|; 53084003 |Bacterial pneumonia (disorder)|) did not significantly improve upon Model One (Model Two: ΔR² = 0.41%, χ²(1) = 3.59, p = .058). Model Three demonstrated that with the addition of the higher specificity SCT concepts, the model could explain significantly more variance in LOS (ΔR² = 4.31%, χ²(1) = 11.51, p < .001). Compared to Model Two, Model Three produced a 50% improvement in the proportion of LOS variance explained. This model contained controls for all predictors from Models One and Two. Therefore, these estimates represent the extra predictive power gained by using more specific SCT concepts.

Table 3.

Estimates and 95% confidence intervals from each of the three models.

	Effect type	Parameter	Model 1	Model 2	Model 3
CCIS (age-adjusted)	Covariate	β =	0.35 [0.23, 0.47]	0.33 [0.21, 0.45]	0.3 [0.17, 0.41]
Sex	Fixed factor [Male = −1, Female = 1]	β =	−0.16 [-0.28, -0.04]	−0.16 [-0.28, -0.03]	−0.14 [-0.26, -0.02]
IMD rank	Covariate	β =	0.21 [0.07, 0.33]	0.21 [0.07, 0.33]	0.2 [0.06, 0.32]
Ethnicity	Random factor	σ =	0.2 [0.06, 0.39]	0.19 [0.04, 0.37]	0.17 [0.02, 0.33]
SCT concept Lower specificity	Fixed factor [Bacterial = −1, viral = 1]	β =	—	−0.12 [-0.23, 0.01]	−0.14 [-0.27, −0.02]
SCT concept Higher specificity	Random factor	σ =	—	—	0.38 [0.29, 0.62]
Patient	Random factor	σ =	1.13 [0.92, 1.34]	1.13[0.92, 1.34]	1.08[0.87, 1.30]
Residual	Residual SD	σ _ε =	1.08 [0.91, 1.30]	1.08 [0.91, 1.31]	1.08 [0.91, 1.31]
AIC			2534.97	2535.8	2523.96
R²			8.25%	8.66%	12.97%
ΔR²			8.25%	0.41%	4.31%
χ²			54.07	3.59	11.51
df			4	1	1
p			<.001	.058	<.001

Note: Table One shows the estimates and their 95% Cis from the three models. For fixed factors and covariates, these are coefficients (denoted β); for random effects, these are standard deviations (denoted σ). Confidence intervals that exclude zero are shown in bold and may be considered “statistically significant”.

Of the well-established predictors (Model One), the most meaningful variable was CCIS (age-adjusted), as shown by its coefficient of 0.35 (corresponding to a 42% increase in length of stay for every one standard deviation (SD) move in CCIS (age-adjusted)) in Table 3. The random effect SD for higher specificity diagnoses (σ = 0.38) corresponds to a 46% increase in LOS for a high specificity diagnosis that is one standard deviation above the mean. This is larger than any of the individual predictors from the other models, demonstrating its comparative importance as a predictor of LOS. All predictors were coded or standardised to represent changes of one SD, so they were statistically comparable.

Discussion

Multiple studies have shown that age, gender, CCIS, ethnicity and IMD rank are significant predictors of LOS (Bahrmann et al., 2019; Cabre et al., 2004; Coevoet et al., 2013; Mathijssen et al., 2016; Yadav et al., 2019). Our results concurred with these conclusions and also showed the importance of using data with high specificity for LOS forecasting. SCT concepts of higher specificity were a more powerful predictor than any individual baseline predictor (Figure 2 and Table 3). Studies have also shown that the aetiology of respiratory illnesses is omitted in healthcare data (Roberts et al., 2016; Roberts and Cheung, 2017). The study results support the findings of these studies and provide further weighting to the need to address the systemic omission of such detail (Table 3). Improvement in data quality characteristics is vital, especially if the data is used in clinical decision support tooling, risk stratification or predictive modelling (Delvaux et al., 2020; Lisk et al., 2019).

The estimate of LOS changed significantly between different subtypes of pneumonia in Model Three (Figure 2). In some cases, this leads to a LOS estimate revision of several whole days, which is of paramount interest to hospital planners and service managers responsible for bed planning. This demonstrates the importance of recording the subtype of pneumonia to support optimal bed management, resource planning and expectation setting with patients.

The model outputs show that considering data quality is essential for predictive modelling. However, not all levels of specificity were as important as others in the study. The lower levels of concept specificity (75570004 |Viral pneumonia (disorder)|) vs (53084003 |Bacterial pneumonia (disorder)|) did not add significant power to the LOS estimate when included in Model Two (ΔR² = 0.41%) (Figure 2). It is important to recognise that although this level of specificity was found not to be relevant for LOS prediction in this study, it is important for other functions such as clinical audit and population health. Enhancing data quality characteristics is always beneficial and desirable as data reuse is high and its applicability wide. Clinicians should be encouraged to record data items at the highest level of specificity wherever possible.

An appropriate level of specificity can be defined as the minimum level needed to address most use for that data. Consideration should be made to all types of healthcare data, including diagnoses, procedures, allergies, histories and events. Clinicians responsible for recording data in health records may be unaware of the whole range of data uses and may never personally reuse the data they capture. Therefore, organisation-wide training and communication on where, why and how to capture high-quality data are paramount. A helpful starting point could be for clinicians and operational managers to work collaboratively to determine the level of data specificity necessary to support core hospital functions and the majority of data needs, such as finance, clinical audit and assurance of evidence-based practice. Continued evaluation of data quality processes should be established to ensure errors and malpractice are rectified (Hickey and Giardino, 2016).

It should be recognised that data entry can be a resource-intensive activity. It is strongly recommended that these processes be as efficient and straightforward as possible to inspire consistent good practice (Batini and Scannapieco, 2016). The results support the findings in current literature, showing that SCT is an effective tool for optimising healthcare data flows (Benson and Grieve, 2016; Lee et al., 2014; Ostropolets et al., 2020; Park et al., 2021).

Improving documentation at the point of entry can reduce the volume and impact of errors on downstream data dependencies, resulting in better care coordination and optimisation. Using EHR-agnostic tools like SCT provide a means to capture high-quality data quickly while preserving desirable structure and standardisation for data processing and application (SNOMED International, 2022c). It may be more time-efficient for a clinician to invest a small amount of time in ensuring they capture the most specific and meaningful terms in front-end systems than it is for them to make sense of low-quality data retrospectively (Batini and Scannapieco, 2016). Where SCT is integrated within front-end systems, generated SCT data is instantly reusable at the point of entry and can support advanced tools such as real-time decision support, safe prescribing and clinical pathway allocation (Delvaux et al., 2020; Lee et al., 2014; Ostropolets et al., 2020; Park et al., 2021). Terminology servers, such as Ontoserver, are effective for scalable implementation and ongoing terminology management (CSIRO, 2022; McBride et al., 2012).

Limitations and future steps

There are opportunities to build on our findings. Only one condition, pneumonia, was assessed, and it is unclear whether the significance and magnitude of the study findings extend to other diseases. Moreover, most primary diagnoses related to pneumonia had no additional modifiers at the study site. As the exact aetiology of these pneumonia cases could not be quickly and reliably ascertained they were excluded from the study. To improve data collection and validity, future studies could link relevant test results to lower specificity pneumonia concepts using post-coordination. For example, an information model could be built linking records for all types of 233604007 |Pneumonia (disorder)| with organisms found through microbiological testing, such as 6415009 |Human respiratory syncytial virus (organism)| (SNOMED International, 2022b). This may increase the cohort sizes and provide a means to include any subtype of pneumonia not already pre-coordinated within an SCT release.The study only assessed aetiology as a means for adding specificity; other means of increasing specificity may be equally or more important. For example, whether a patient had a case of bronchopneumonia or lobar pneumonia.

This study was limited to one centre, so the results do not necessarily reflect the patient case-mix of the other UK or international centres. Extending the study to include a range of conditions across multiple sites would give a more reliable indicator of the importance of data specificity in LOS prediction. The estimates for the higher specificity data items had large confidence intervals due to relatively low sample sizes within each (Figure 2). Higher volumes of data will result in more accurate estimates for all parameters with smaller confidence intervals. Broader and faster national adoption of SCT is also encouraged to expand the number of data items available for study. Additional variables could be included to assess the relevance of data-item specificity further. Current literature has shown other data points to be relevant, including admission observations, admission time or specific pre-existing comorbidities (Abd-Elrazek et al., 2021; Daghistani et al., 2019; Lisk et al., 2019). SCT concept specificity may need consideration when predicting other metrics, including occurrence of accident and emergency attendances, adverse events and healthcare inequalities. These are critical areas for investigation.

It is a clinical judgement as to whether microbiological testing is required. As such, it is not always completed for every pneumonia presentation, and where it is conducted, results may return inconclusive (Lee et al., 2019). The findings for short stays (<3 days) may have less precision when compared to more prolonged admissions, as there is less time for testing to be completed and repeated (if necessary). These cases may have been more likely to have an SCT concept of greater specificity recorded and subsequently included within the study. To overcome this limitation, future studies could include testing on or soon after admission for all suspected or confirmed cases of pneumonia. Such practice aligns with growing evidence that rapidly determining the microbial aetiology of pneumonia is imperative to ensure effective antibiotic administration (Murdoch et al., 2009).

Conclusion

Specificity is an essential characteristic of high-quality data. SCT data with higher specificity explained more variance in LOS than any individual predictor, including gender, ethnicity, IMD rank and CCIS (age-adjusted). SCT was shown to be an effective tool for healthcare analytics, and as such scalable implementation and use are important. Clinicians are encouraged to use SCT to capture healthcare data items with the highest specificity possible. Robust data quality training and assurance processes should accompany all healthcare data flows.

Supplemental Material

Supplemental Material - The importance of SNOMED CT concept specificity in healthcare analytics

Supplemental Material for The importance of SNOMED CT concept specificity in healthcare analytics by Luke Roberts, Sadie Lanes, Oliver Peatman, and Phil Assheton in Health Information Management Journal

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This paper would not have been possible without the exceptional support of the clinical analytics team at Guy’s and St Thomas’ NHS Foundation Trust, London England.

ORCID iDs

Luke Roberts

Oliver Peatman

Supplemental material

Supplemental material for this article is available online.

References

Abd-Elrazek

Eltahawi

Abd Elaziz

, et al. (2021) Predicting length of stay in hospitals intensive care unit using general admission features. Ain Shams Engineering Journal 12(4): 3691–3702. DOI: 10.1016/j.asej.2021.02.018.

Alonso

Santos

Pinto

, et al. (2020a) Health records as the basis of clinical coding: Is the quality adequate? A qualitative study of medical coders’ perceptions. Health Information Management Journal, 49: 28–37. DOI: 10.1177/1833358319826351.

Alonso

Santos

Pinto

, et al. (2020b) Problems and Barriers during the Process of Clinical Coding: a Focus Group Study of Coders’ Perceptions. Journal of Medical Systems, 44: 1–8. DOI: 10.1007/S10916-020-1532-X.

Bahrmann

Benner

Christ

, et al. (2019) The Charlson Comorbidity and Barthel Index predict length of hospital stay, mortality, cardiovascular mortality and rehospitalization in unselected older patients admitted to the emergency department. Aging Clinical and Experimental Research, 31: 1233–1242. DOI: 10.1007/s40520-018-1067-x.

Baniasadi

Kahnouji

Davaridolatabadi

, et al. (2019) Factors affecting length of stay in Children Hospital in Southern Iran. BMC Health Services Research, 19: 1. DOI: 10.1186/s12913-019-4799-1.

Bates

Mächler

Bolker

, et al. (2015) Fitting Linear Mixed-Effects Models Using lme4. Journal of Statistical Software 67(1): 11. DOI: 10.18637/jss.v067.i01.

Batini

Scannapieco

(2016) Data and Information Quality: Dimensions, Priniciples, and Techniques. In Data-Centric Systems and Applications. 1st ed. Cham: Springer International Publishing. DOI: 10.1007/978-3-319-24106-7.

Benson

Grieve

(2016) Principles of Health Interoperability. 3rd ed. Cham: Health Information Technology Standards. Springer. DOI: 10.1007/978-3-319-30370-3.

Bruce

Smith

(2020) Length of stay among multi-ethnic psychiatric inpatients in the United Kingdom. Comprehensive Psychiatry 102: 152201. DOI: 10.1016/j.comppsych.2020.152201.

10.

Cabre

Bolivar

Pera

, et al. (2004) Factors influencing length of hospital stay in community-acquired pneumonia: a study in 27 community hospitals. Epidemiology and Infection 132(5): 821–829. DOI: 10.1017/S0950268804002651.

11.

Chalmers

Campling

Ellsbury

, et al. (2017) Community-acquired pneumonia in the United Kingdom: a call to action. Pneumonia 9(1): 15. DOI: 10.1186/s41479-017-0039-9.

12.

Charlson

Szatrowski

Peterson

, et al. (1994) Validation of a combined comorbidity index. Journal of Clinical Epidemiology 47(11): 1245–1251. DOI: 10.1016/0895-4356(94)90129-5.

13.

Cirillo

Nunez Carpintero

Valencia

(2021) Artificial intelligence in cancer research: learning at different levels of data granularity. Molecular Oncology 15(4): 817–829. DOI: 10.1002/1878-0261.12920.

14.

Coevoet

Fresson

Vieux

, et al. (2013) Socioeconomic deprivation and hospital length of stay: A new approach using area-based socioeconomic indicators in multilevel models. Medical Care 51(6): 548–554. DOI: 10.1097/MLR.0b013e3182928f84.

15.

CSIRO (2022) Ontoserver: Shrimp : SNOMED ECL Builder. Available at: https://ontoserver.csiro.au/shrimp/ecl/?fhir=https://ontology.nhs.uk/authoring/fhir (accessed 10 February 2022).

16.

Daghistani

Elshawi

Sakr

, et al. (2019) Predictors of in-hospital length of stay among cardiac patients: A machine learning approach. International Journal of Cardiology 288: 140–147. DOI: 10.1016/j.ijcard.2019.01.046.

17.

Davoudi

Dooling

Glondys

, et al. (2015) Data Quality Management Model 2015 Update-Retired. Journal of AHIMA 86(10). Available at: https://library.ahima.org/PB/DataQualityModel#.YemFg477SUl (accessed 20 January 2022).

18.

Delvaux

Vaes

Aertgeerts

, et al. (2020) Coding Systems for Clinical Decision Support: Theoretical and Real-World Comparative Analysis. JMIR Formative Research 4(10): 11. JMIR Form Res. DOI: 10.2196/16094.

19.

Department for Levelling UpH and C and Ministry of Housing C and LG (2020) English Indices of Deprivation. Available at: https://www.gov.uk/government/collections/english-indices-of-deprivation (accessed 23 June 2022).

20.

Department of Health and Social Care (2018) The future of healthcare: our vision for digital, data and technology in health and care. UK Government. Available at: https://www.gov.uk/government/publications/the-future-of-healthcare-our-vision-for-digital-data-and-technology-in-health-and-care/the-future-of-healthcare-our-vision-for-digital-data-and-technology-in-health-and-care (accessed 20 January 2022).

21.

Georgalas

Tolley

Narula

(2014) Tonsillitis. BMJ Clinical Evidence 2014. 1: 1

22.

Guy’s and St Thomas’ NHS Foundation Trust (2022) Our Organisation. Available at: https://www.guysandstthomas.nhs.uk/about-us/our-organisation (accessed 23 June 2022).

23.

Hickey

Giardino

(2016). In: Hickey

J v.

Brosnan

(eds), Evaluation of Health Care Quality for DNPs. 3rd ed. New York, NY: Springer Publishing Company. DOI: 10.1891/9780826131584.

24.

Lee

Cornet

Lau

, et al. (2013) A survey of SNOMED CT implementations. Journal of Biomedical Informatics 46(1): 87–96. DOI: 10.1016/j.jbi.2012.09.006.

25.

Lee

de Keizer

Lau

, et al. (2014) Literature review of SNOMED CT use. Journal of the American Medical Informatics Association, 21. e11–e19. DOI: 10.1136/AMIAJNL-2013-001636.

26.

Lee

Ruan

S-Y

Pan

S-C

, et al. (2019) Performance of a multiplex PCR pneumonia panel for the identification of respiratory pathogens and the main determinants of resistance from the lower respiratory tract specimens of adult patients in intensive care units. Journal of Microbiology, Immunology and Infection 52(6): 920–928. DOI: 10.1016/j.jmii.2019.10.009.

27.

Lequertier

Wang

Fondrevelle

, et al. (2021) Hospital length of stay prediction methods: A Systematic Review. Medical Care, 59: 929–938. DOI: 10.1097/MLR.0000000000001596.

28.

Lisk

Uddin

Parbhoo

, et al. (2019) Predictive model of length of stay in hospital among older patients. Aging Clinical and Experimental Research 31(7): 993–999. DOI: 10.1007/s40520-018-1033-7.

29.

Mathijssen

NMC

Verburg

van Leeuwen

CCG

, et al. (2016) Factors influencing length of hospital stay after primary total knee arthroplasty in a fast-track setting. Knee Surgery, Sports Traumatology, Arthroscopy 24(8): 2692–2696. DOI: 10.1007/s00167-015-3932-x.

30.

McBride

Lawley

Leroux

, et al. (2012) Using Australian Medicines Terminology (AMT) and SNOMED CT-AU to better support clinical research. Studies in Health Technology and Informatics, 178: 144. DOI: 10.3233/978-1-61499-078-9-144.

31.

Millett

ERC

de Stavola

Quint

, et al. (2015) Risk factors for hospital admission in the 28 days following a community-acquired pneumonia diagnosis in older adults, and their contribution to increasing hospitalisation rates over time: a cohort study. BMJ Open 5(12): e008737. DOI: 10.1136/bmjopen-2015-008737.

32.

Murdoch

O’Brien

Scott

JAG

, et al. (2009) Breathing New Life into Pneumonia Diagnostics. Journal of Clinical Microbiology 47(11): 3405–3408. DOI: 10.1128/JCM.01685-09.

33.

Nakagawa

Schielzeth

(2013) A general and simple method for obtaining R ² from generalized linear mixed-effects models. Methods in Ecology and Evolution 4(2): 133–142. DOI: 10.1111/j.2041-210x.2012.00261.x.

34.

National Information Board (2014) Personalised Health and Care 2020 Using Data and Technology to Transform Outcomes for Patients and Citizens A Framework for Action. UK Government. November 2014.

35.

NHS Digital (2020) SCCI0034: SNOMED CT. Available at: https://digital.nhs.uk/data-and-information/information-standards/information-standards-and-data-collections-including-extractions/publications-and-notifications/standards-and-collections/scci0034-snomed-ct.(accessed 23 June 2022)

36.

NHS Digital (2022a) National data opt-out. Available at: https://digital.nhs.uk/services/national-data-opt-out(accessed21 January 2022).

37.

NHS Digital (2022b) The NHS Digital Terminology Server. Available at: https://digital.nhs.uk/services/terminology-servers (accessed 23 June 2022).

38.

Ono

Tamai

Takeuchi

et al. (2010) Predictors of length of stay in a ward for demented elderly: gender differences. Psychogeriatrics 10(3): 153–159. doi: 10.1111/j.1479-8301.2010.00328.x.

39.

Ostropolets

Reich

Ryan

, et al. (2020) Characterizing database granularity using SNOMED-CT hierarchy. In AMIA Annual Symposium Proceedings 2020. American Medical Informatics Association: 983

40.

Park

H-A

S-J

Jung

(2021) Strategies for Adopting and Implementing SNOMED CT in Korea. Healthcare Informatics Research 27(1): 3–10. DOI: 10.4258/hir.2021.27.1.3.

41.

Roberts

Cheung

(2017) The aetiology of infections: The significance for hospital management. British Journal of Health Care Management, 23: 271–280

42.

Roberts

Timmis

Nagrecha

, et al. (2016) Capturing causative agents in clinical coding. British Journal of Health Care Management, 22: 461–468 DOI: 10.12968/bjhc.2016.22.9.461.

43.

Roberts

Lanes

Kyte

, et al. (2022) Acute pain assessments and records: a pilot study of digital transformation. British Journal of Nursing 31(10): 541–548. DOI: 10.12968/bjon.2022.31.10.541.

44.

SNOMED International (2020) Frequently Asked Questions: Regulation. Available at: https://confluence.ihtsdotools.org/display/DOCNRCFAQ/3.+Regulation (accessed 9 March 2022).

45.

SNOMED International (2022a) Case Studies. Available at: https://www.snomed.org/snomed-ct/snomed-in-action/search-results?region=EMEA (accessed 23 June 2022).

46.

SNOMED International (2022b). SNOMED CT Browser - International Edition. Available at: https://browser.ihtsdotools.org/? (accessed 9 March 2022).

47.

SNOMED International (2022c) SNOMED CT Starter Guide. Available at: https://confluence.ihtsdotools.org/display/DOCSTART/SNOMED+CT+Starter+Guide (accessed 21 January 2022).

48.

Yadav

Unal

Armstrong

, et al. (2019) P806 Impact of patient age, gender and season of admission on length of stay in hospital for acute inflammatory bowel disease admissions. Journal of Crohn’s and Colitis 13(Supplement_1): S525–S525. DOI: 10.1093/ecco-jcc/jjy222.930.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.31 MB