(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .
Hospital-wide natural language processing summarising the health data of 1 million patients [1]
['Daniel M. Bean', 'Department Of Biostatistics', 'Health Informatics', 'Institute Of Psychiatry', 'Psychology', 'Neuroscience', 'King S College London', 'London', 'United Kingdom', 'Health Data Research Uk London']
Date: 2023-07
Electronic health records (EHRs) represent a major repository of real world clinical trajectories, interventions and outcomes. While modern enterprise EHR’s try to capture data in structured standardised formats, a significant bulk of the available information captured in the EHR is still recorded only in unstructured text format and can only be transformed into structured codes by manual processes. Recently, Natural Language Processing (NLP) algorithms have reached a level of performance suitable for large scale and accurate information extraction from clinical text. Here we describe the application of open-source named-entity-recognition and linkage (NER+L) methods (CogStack, MedCAT) to the entire text content of a large UK hospital trust (King’s College Hospital, London). The resulting dataset contains 157M SNOMED concepts generated from 9.5M documents for 1.07M patients over a period of 9 years. We present a summary of prevalence and disease onset as well as a patient embedding that captures major comorbidity patterns at scale. NLP has the potential to transform the health data lifecycle, through large-scale automation of a traditionally manual task.
Clinical notes and letters are still the main way that medical information is recorded and shared between clinical staff. This means that for research we need methods that can cope with text data, which is typically far more challenging than “structured” data like diagnosis codes or test results. In this study we apply a state of the art clinical text processing model to analyse almost 10 years worth of text data from a large London hospital, covering over 1 million patients. We are able to find patterns of disease burden, onset and co-occurrence purely in text data. This result strongly supports the use of clinical text data in research and provides a summary of the scale and nature of clinical text to other researchers.
Competing interests: I have read the journal’s policy and the authors of this manuscript have the following competing interests: JTT has previously received research grant support from Innovate UK, NHSX, Office of Life Sciences, NIHR, Health Data Research UK, Bristol-Meyers-Squibb and Pfizer; has received honorarium from Bayer, Bristol-Meyers-Squibb and Goldman Sachs; holds stock in Amazon, Alphabet, Nvidia; and receives royalties from Wiley-Blackwell Publishing. DMB has received research funding from Pfizer. RJBD, ZK, AS declare that no competing interests exist.
Funding: The project has received funding support from Innovate UK, NHS AI Lab, Office of Life Sciences, Health Data Research UK, NIHR Maudsley Biomedical Research Centre and NIHR Applied Research Centre South London. DMB is funded by Health Data Research UK and NHS AI Lab. RJBD is supported by the following: (1) NIHR Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King’s College London, London, UK; (2) Health Data Research UK, which is funded by the UK Medical Research Council, Engineering and Physical Sciences Research Council, Economic and Social Research Council, Department of Health and Social Care (England), Chief Scientist Office of the Scottish Government Health and Social Care Directorates, Health and Social Care Research and Development Division (Welsh Government), Public Health Agency (Northern Ireland), British Heart Foundation and Wellcome Trust; (3) The BigData@Heart Consortium, funded by the Innovative Medicines Initiative-2 Joint Undertaking under grant agreement No. 116074. This Joint Undertaking receives support from the European Union’s Horizon 2020 research and innovation programme and EFPIA; it is chaired by DE Grobbee and SD Anker, partnering with 20 academic and industry partners and ESC; (4) the National Institute for Health Research University College London Hospitals Biomedical Research Centre; (5) the National Institute for Health Research (NIHR) Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King’s College London; (6) the UK Research and Innovation London Medical Imaging & Artificial Intelligence Centre for Value Based Healthcare; (7) the National Institute for Health Research (NIHR) Applied Research Collaboration South London (NIHR ARC South London) at King’s College Hospital NHS Foundation Trust; (8) NHS AI Lab. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Data Availability: The full patient-level dataset contains sensitive and potentially re-identifiable data and cannot currently be made available directly. To support data availability, we have registered the dataset on the Health Data Research UK Innovation Gateway online (
https://web.www.healthdatagateway.org/dataset/4e8d4fed-69d6-402c-bd0a-163c23d6b0ee ). This provides a thorough description (with the highest tier metadata score awarded by the platform(35)) and procedure for access of summary or aggregated data (to fully eliminate re-identification risk).
In this paper we present the first descriptive summary of the entire text record of a large UK secondary and tertiary healthcare system in London, King’s College Hospital NHS Foundation Trust over a period of about 9 years. To our knowledge this is the first study of a large-scale EHR dataset derived from NLP, although there are several other descriptive analyses of large-scale structured EHR data (Kuan et al. 2019; Thygesen et al. 2022; Kuan et al. 2023). Compared to structured data, the free-text portion of the EHR captures a more detailed clinical narrative. The description of this data provides three useful resources:
Natural Language Processing (NLP) combined with rich clinical terminologies such as SNOMED have the potential to automate a large portion of the ‘structure and standardise’ process to make the full clinical record accessible to computational analysis [ 7 – 9 ]. Previous attempts have focused on specific cohorts (e.g. critical care patients only [ 10 ], patients with a certain disease only [ 11 – 13 ], discharge letters only [ 14 ]). Doing this across a whole hospital’s record has not previously been attempted, and produces the opportunity to automate a laborious manual process for healthcare delivery, and also to enrich any structured registries or databases (like HES [ 15 ], SUS [ 16 ], CPRD [ 17 ], Caliber [ 18 ], CVD-Covid-UK [ 19 ]) with greater phenotypic and narrative expressiveness. Any downstream data-dependent activity, including population health and research, or trial recruitment [ 20 ], would potentially benefit.
In conventional healthcare workflows, both structured and unstructured aspects of EHR’s are read by business intelligence staff and translated into standardised codes (termed ’clinical coders’) for submissions into datasets. Structured data can be analysed at a regional or national level to gain powerful insights into clinical trajectories at scale [ 5 , 6 ]. This largely manual process uses the ICD10, OPCS ontologies and follows rules around conciseness. Due to the laborious nature of this process and lack of an automation-assisted process, most organisations only perform this ’structuring and standardising’ process on inpatient episodes and the text generated from the large proportion of outpatient activity are ignored. This ’lacune’ means that certain populations with conditions that do not result in hospitalisations (or where clinical pathway transformations migrate to ambulatory or outpatients routes) would be under-represented systematically; dependency on only manually-derived coded data potentially incorporates a hidden ‘inclusion bias’ in many datasets.
Electronic Health Records (EHRs) are now widely deployed, and in many cases these electronic systems have accumulated a considerable history of clinical data. Each clinical site therefore represents a potentially significant data resource. There is considerable structured coding of clinical events and related results, and the structured data capture is highly targeted to specific purposes (primarily billing or reporting). Such structured diagnosis lists, problem lists or test lists often only partially capture the full clinical picture of a patient as the primary means of clinical communication and documentation is in the form of free text letters, notes and reports [ 1 ]. Most analytical quantitative research have focused on the structured elements only as the unstructured free text recorded in EHRs have traditionally been difficult to access and analyse [ 2 – 4 ]
The colour assignment in Fig 4 should not be taken to mean categorical differences between two neighbouring clusters: for example we note two neighbouring clusters representing asthma patients with and without hypertension, and the cluster separation being an artefact of the clustering algorithm (e.g. patients in the asthma-hypertension group being simply on the same asthma cluster but on a spectrum). Although the clustering algorithm assigns hard borders between clusters, it is important that the embedding is a continuous space and we are not suggesting there are hard divisions between these patients.
As we are only considering disorders, we are not attempting to uncover a globally optimal set of patient clusters. Instead we primarily use the clustering algorithm to sample contiguous regions of the patient embedding space which we can test for enrichment.
A sample of 100,000 patients was embedded based on normalised annotation counts for all SNOMED disorder codes detected in at least 1000 patients at KCH. Colour indicates cluster membership (50 clusters), text boxes indicate major disorders in the indicated cluster with the percent of patients in brackets. Where a particular disease was predominant within a broader category it is also shown within brackets. The grey box on the left is a detailed view of the region indicated with a dashed grey border in the main plot area. COPD = chronic obstructive pulmonary disease, UTI = urinary tract infection. This visualisation is available in interactive form online [ 22 ].
Each patient in the dataset can be represented as a vector of NLP annotation counts. These counts capture a complex interrelationship of comorbidity, treatment and outcome. To qualitatively capture these comorbidity patterns, we generated a low-dimensional embedding from the input vectors from a random sample of 100k patients and used agglomerative clustering with ward linkage to sample regions of this embedding space. The sample was not significantly different to the remaining patients in the dataset in their age (t-test p = 0.45) or sex (fisher exact test p = 0.90) distribution. As shown in Fig 4 , the resulting embedding with 50 clusters contains regions of patients that are strongly associated with major diagnoses. For 72% of clusters, at least one SNOMED disorder was present in over 50% of cluster patients. For 48% of clusters, the most common disorder was present in at least 75% of the cluster patients. The annotations in Fig 4 include propagating counts through the SNOMED ontology (see Methods ). An interactive visualisation of the embedding in Fig 4 is available online [ 22 ].
Each disease is represented by a number of specific SNOMED codes. Ages are shown as mean with standard deviation in brackets. Prevalences are shown as percentages with counts in brackets. Age distributions were compared with a t-test and prevalences were compared with fisher’s exact test.
Analysis of these diseases by sex finds many significant differences in both mean age at first detection and overall prevalence ( Table 3 ). A particularly large, significant difference in first detection is noted for psychosis (46.12 for males, 50.87 for females, p<0.01). Although we do detect a younger first detection age for DM1 vs DM2 in both male and female patients, the age distribution for DM1 is not in line with expectations.
Fig 3 shows the age profile of the various diagnosis codes are compatible with expectations with diseases of young adults like Multiple Sclerosis (MS), Psychosis and Inflammatory Bowel Disease, while diseases of later ages being degenerative (dementia or Parkinson’s Disease) or related to end-organ failure (heart failure). These age profiles suggest that large-scale NLP extraction from clinical documents produces datasets with similar characteristics to standardised national datasets.
In addition to the presence of a code, we can also analyse the age at first detection of a code in the record. Note that this will not always be age at diagnosis. 14 major conditions were manually identified to provide a breadth of clinical specialties and expected onset.
Across all conditions for which we could implement the same definition as used for QoF in our data we found a significant (p<0.001) difference in prevalence compared to primary care estimates for England and London. There were only 2 conditions for which we found a lower prevalence than the London estimate; CKD and Depression. In particular we find a rate of CKD that is less than half the London GP prevalence. 3782 patients in our data meet the QoF definition (CKD grade 3a to 5), and an additional 2546 patients for which we detected CKD but not a grade, so they cannot be included but could potentially have CKD grade 3a-5. Mild CKD may not have been explicitly typed by clinicians into letters or notes, but this under-detection of CKD would be addressable using structured EHR data (i.e. eGFR blood result) as well as to map out the CKD to the different grades.
Prevalence estimates were calculated for all conditions for which the Quality and Outcomes Framework definitions could be directly mapped to our data including all admissions from 01 April 2018–31 March 2019. Conditions were defined using the NHS Digital business rules version 41 as used in the 2018/19 Quality and Outcomes Framework report. All pairwise comparisons between our data and the England or London estimates were significant at p<0.001 level following fisher’s exact test and Bonferroni correction for multiple testing.
We used the NHS England business rules (which define specific conditions using SNOMED codes [ 23 ]) and the NHS England Quality and Outcomes Framework (QoF) data to calculate prevalence of disorders included in the 2018/19 QoF report [ 24 ] according to the same definitions (i.e. sets of included codes) for each condition. The prevalence estimates for our data and the QoF results for England and London (for national and regional context) are shown in Table 2 .
We found the overall most prevalent concept was hypertension (14.65%) which was also the most prevalent disorder for both male and female patients. Note that these prevalence estimates are for single SNOMED codes (see “Comparison to National Data” prevalence estimates using code sets). However, the estimates are consistent with the QoF data for 2018–19 which found a prevalence of 13.96% for hypertension in England. Similarly, we estimate prevalence of asthma at 5.69% and the QoF estimate is 6.05% for England. However there are discrepancies, such as for depression (5.55% in our data, 10.74% in QoF for England, 7.65% for QoF in London).
1904 SNOMED codes (11.8% of codes detected) had a significant difference in prevalence between male and female patients (t-test p < 0.05 after correction for multiple comparisons). Table 1 shows the top 10 most prevalent disorders overall and by gender. Fig 2 shows the prevalence of all 14 unique disorders from Table 1 across all groups. All differences in prevalence in Fig 2 and Table 1 were significant at p < 0.001 level. Many of the other differences in prevalence are expected sex-specific conditions.
Fig 1 shows the breakdown of annotations by semantic type and meta annotation class. Most annotations were to SNOMED disorder, finding, substance and procedure classes, i.e. diseases, symptoms and treatments. Interactive treemap visualisations for the top 100 most common SNOMED codes detected for semantic types Finding, Disorder and Substance are available online [ 22 ]. The vast majority of annotations were positive (85.9%), referred to the patient (93.1%) and are current/recent in time (92.9%). We therefore focus on this subset of 123M annotations (78.3%) where all three criteria are true in the remaining analysis.
We extracted data from 2011-01-01 to 2019-12-31 for all patients aged 18–100 years at the time of admission. The dataset includes 1.07M patients and 212M separate text notes. 157M NLP annotations to SNOMED concepts were generated for those notes. We found that the scale of data available has increased over time, partially due to an increase in the number of patients per year but, importantly, we also captured more data per patient over time. The number of patients in the dataset increased from ~165 to ~369k patients per year while the median annotations per patient per year increased from a median of 14 to 25. The increase of patient numbers is due to the merger of KCH with a suburban secondary care hospital (Princess Royal University Hospital, PRUH) in 2015, with subsequent data incorporation.
Discussion
We present a summary of the text records for over 1M patients over almost a decade. We find that the rate of data capture per patient is increasing over time. The dataset is generated entirely by NLP applied to clinical text and captures major trends in disease prevalence and age of onset.
A number of significant differences in prevalence between male and female patients were observed. Depressive disorder was significantly more prevalent in female patients (5.90% vs 5.12%), as was asthma (6.44% vs 4.78%). In both cases the difference is consistent with expectations. We also find a significantly higher prevalence of dementia in female patients (2.47% vs 2.3%) and a later age at first detection (77.73 years vs 80.15 years). The greater prevalence of dementia in female patients is well established. The age at first detection in our record is not necessarily the same as age of diagnosis, as a new patient could arrive with a known history of dementia. However, there is evidence that female dementia patients tend to be diagnosed later in life.
Prevalence estimates calculated for our data reflect a specific clinical context of secondary care inpatients. This cohort should be expected to have different prevalence of conditions from primary care (higher prevalence of most conditions) and also local differences in sociodemographic factors. Although we demonstrate many areas of strong agreement with expected trends from national data, there are discrepancies. A significant artefactual error is noted, eg: Type 1 Diabetes Mellitus which is known to have a young adult onset but this dataset captures a lot of middle-age and elderly onset cases. This is likely due to the NLP misattributing late onset Type II diabetes mellitus patients who are commencing on insulin—a vocabulary error of source document which is highly contextual to the specific snomed code as Type 1 and Type 2 due to the deprecated outdated concepts of insulin-dependent diabetes mellitus (IDDM) and non-insulin-dependent diabetes mellitus (NIDDM). This is correctable by setting the NER+L to recognise “IDDM” as an ambiguous token and not mapping it to “Type 1 Diabetes Mellitus (SCTID 46635009)”. The high prevalence of cysts (rank 5 for female patients, 4.99%) is likely due to most mentions of cysts being annotated to the generic concept rather than a more specific term.
The strength of our dataset is its application to the in-hospital, secondary and tertiary care setting. We demonstrate that the large-scale analysis of NLP phenotypes can identify clusters of patients with differing major diseases, and find evidence that the embedding space captures clinical spectrums in some areas (e.g. patients with and without a major comorbidity). This suggests that outcomes of interest could be overlaid on the embedding space to find associations. The embedding is also potentially a powerful tool to identify data-driven cohorts of patients from high-dimensional data.
To our knowledge there is no other large scale EHR dataset that is derived from NLP. Related resources in the UK include Hospital Episode Statistics (HES), Clinical Practice Research Datalink (CPRD) and CALIBER. These sources all derive from structured data, primarily diagnosis codes assigned during a primary or secondary care episode. These codes have the advantage that they are manually assigned by trained experts, but the disadvantage is that they are collected primarily for billing/commissioning ‘business’ purposes and are not necessarily intended to capture all known comorbidities, procedures or medications for the patient. Structured codes are also not error-free [25,26]. The free text narrative, in contrast, is typically much more expressive and detailed as it is designed to create a full record for other clinical staff to rely on (including clinical coding teams).
The challenges of secondary use of real-world data, particularly text data, are not only technical [27]; ethical and legal processes must also be in place. King’s College Hospital operates a patient expert-led oversight committee, similar to the model in place at pioneering sites such as the South London and Maudsley CRIS system. Performant NLP is a necessary step to unlocking the research potential of EHRs, but it is not sufficient without similar supporting ethical and legal infrastructure.
[END]
---
[1] Url:
https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000218
Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.
via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/
