开发用于评估骨髓增生性肿瘤心血管风险的自然语言处理管道。

IF 7.6 2区 医学 Q1 HEMATOLOGY
HemaSphere Pub Date : 2024-08-08 DOI:10.1002/hem3.143
Andrea Duminuco, Joshua Au Yeung, Raj Vaghela, Sukhraj Virdee, Claire Woodley, Susan Asirvatham, Natalia Curto-Garcia, Priya Sriskandarajah, Jennifer O'Sullivan, Hugues de Lavallade, Deepti Radia, Shahram Kordasti, Giuseppe Palumbo, Claire Harrison, Patrick Harrington
{"title":"开发用于评估骨髓增生性肿瘤心血管风险的自然语言处理管道。","authors":"Andrea Duminuco,&nbsp;Joshua Au Yeung,&nbsp;Raj Vaghela,&nbsp;Sukhraj Virdee,&nbsp;Claire Woodley,&nbsp;Susan Asirvatham,&nbsp;Natalia Curto-Garcia,&nbsp;Priya Sriskandarajah,&nbsp;Jennifer O'Sullivan,&nbsp;Hugues de Lavallade,&nbsp;Deepti Radia,&nbsp;Shahram Kordasti,&nbsp;Giuseppe Palumbo,&nbsp;Claire Harrison,&nbsp;Patrick Harrington","doi":"10.1002/hem3.143","DOIUrl":null,"url":null,"abstract":"<p>A central feature of myeloproliferative neoplasms (MPN) is an increased risk of cardiovascular thrombotic complications, and this is the primary determinant for the introduction of cytoreductive therapy.<span><sup>1</sup></span> The landmark ECLAP study in polycythemia vera (PV) patients, showed cardiovascular mortality accounted for 45% of all deaths, with a thrombosis incidence rate of 1.7/100 person/year and a cumulative incidence of 4.5% over a median follow-up of 2.8 years.<span><sup>2</sup></span></p><p>Natural language processing (NLP) is a branch of machine learning involving computational interpretation and analysis of human language. CogStack (https://github.com/CogStack), is an open-source software ecosystem, that retrieves structured and unstructured components of electronic health records (EHR). The Medical Concept Annotation Toolkit (MedCAT), the NLP component of CogStack, structures clinical free text by disambiguating and capturing synonyms, acronyms, and contextual details, such as negation, subject, and grammatical tense, and mapping text to medical Systematized Nomenclature of Medicine–Clinical Terms (SNOMED-CT) concepts. This technique is known as “named entity recognition and linkage” (NER+L). MedCAT has previously been used and validated in many studies to structure EHR data across a range of medical specialties for auditing, observational studies, de-identifying patient records, operational insights, disease modeling, and prediction.<span><sup>3-8</sup></span></p><p>We employed our NLP pipeline, Cogstack, and MedCAT, to determine the prevalence and impact of cardiovascular risk factors upon thrombotic events during follow-up. We used Cogstack to retrieve outpatient hematology clinic letters and hematology discharge letters. MedCAT was then used for NER+L of relevant clinical free-text to respective SNOMED-CT codes that were determined by two hematology specialists. The base MedCAT model was trained unsupervised on &gt;18 million EHR documents, and this was further fine-tuned using a 80:20 train:test split with 600 clinician-annotated MPN-specific documents. Total SNOMED-CT code counts were aggregated and grouped by individual patient, a unique threshold count was then applied to “infer” presence of the respective SNOMED code. In this process, hematology specialists read through clinical documents and manually highlight correct words or phrases detected by MedCAT that correspond to the SNOMED concept of interest.</p><p>We deploy a two-step validation process that has been well described.<span><sup>3, 8</sup></span> The first is to evaluate and validate the NER model performance on a document level demonstrating how accurately MedCAT is able to identify the medical concepts of interest. This involves hematology specialists annotating medical concepts and comparing this to the model NER outputs (Supporting Information S1: Table 1). The second step involves manual validation by creating a gold-standard real-world dataset. Two hematology specialists were randomly assigned to review a subsample of the patient sample's clinical notes (<i>n</i> = 112 [20%] in ET cohort, and <i>n</i> = 60 [17%] in the PV cohort) and, taking into account the entire clinical history, state whether the patient had presence or absence of the selected SNOMED concepts (Supporting Information S1: Tables 2 and 3). Finally, using the manual-validated data set, a threshold optimizer was used to find optimal concept counts for real-world F1 inference. The manual validation is crucial because whilst ML models can mislabel concepts on a document level, the key factor is how well a model performs in the real-world at recognizing the presence or absence of a clinical entity on a patient level.</p><p>The selected SNOMED concepts were cardiovascular risk factors, including hypertension (HTN), hypercholesterolemia (HC), diabetes mellitus (DM), smoker status, and obesity. We also assessed for cardiovascular events, including portal vein thrombosis (PVT), deep vein thrombosis (DVT), pulmonary embolism (PE), myocardial infarction (MI), stroke/cerebrovascular accident (CVA), cerebral sinus thrombosis (CST), and thrombosis not otherwise specified (NOS). A schematic representation of the process is represented in Figure 1A.</p><p>Data from 360 PV and 560 ET patients, reviewed at Guys' and St Thomas NHS Foundation Trust (GSTT) for at least one visit between January 2005 and April 2023, were evaluated (Supporting Information S1: Table 4). A total of 12905 documents from 560 ET patients (median 20 per patient, interquartile range [IQR], 8–34), and 11250 from 360 PV patients, (median 27 per patient, IQR, 11–47), were reviewed. In the manual validation data set (<i>n</i> = 112 patients for ET and <i>n</i> = 60 for PV), MedCAT achieved excellent real-world F1 scores with regards to specificity and sensitivity (Supporting Information S1: Tables 2 and 3).</p><p>For the ET cohort, using a threshold of &gt;2 mentions to define a positive population, HTN was identified in 21.3% (119) of patients, DM in 4.6% (26), MI in 3.6% (20), CVA in 7.7% (43), NOS thrombosis in 8% (45), DVT in 1.4% (8), PE in 1.8% (10), PVT in 1.3% (7) and positive smoking status in 6.6% (37) (Supporting Information S1: Figure 1A,B). HC was identified in 9.6% (54) using a threshold &gt;1. 52% (56) of patients with HC and 69.2% (18) of those with DM also had HTN. Obesity was not identified in any patients using this approach. Considering overall venous thromboembolic (VTE) occurrence, 11.6% (65) of patients reported an event. Of patients with CVA/MI, 58.1% (43)/55% (11) had this event pre- or at diagnosis, and 30.2% (13)/10% (2) while receiving cytoreductive therapy.</p><p>For the PV cohort, using a threshold of &gt;2 mentions to define the presence of the condition, HTN was identified in 23.1% (83) of cases, DM in 5.6% (20), MI in 3.1% (11), NOS thrombosis in 19.4% (70), DVT in 2.8% (10), PE in 2.8% (10), and PVT in 5% (18) with only one case of cerebral venous sinus thrombosis identified (0.3%) (Supporting Information S1: Figure 2A,B). Positive smoking habit was reported in 5.6% (20) of patients. Overall VTE was reported in 23.3% (84) of cases. Using a threshold of 1 mention, HC and CVA were detected in 7.5% (27) and 14.2% (51) patients, respectively. In keeping with the ET cohort, obesity was not identified in any patients. CVA/MI events occurred in 58.8% (30)/54.5 (6) cases prior to or at diagnosis and 23.5% (12)/9.1 (1) during cytoreductive treatment. Patients affected by DM suffered from HTN in 75% of cases (15), 66.7% (18) of HC patients had concomitant diagnosis of HTN, while 1.7% of the total had HTN, HC, and DM together.</p><p>TE were significantly higher in PV patients in the case of CVA (<i>p</i> = 0.002), PVT (<i>p</i> &lt; 0.001), venous thromboembolism (VTE, referring to any form of venous thrombotic event in an anatomical region that is not otherwise mentioned, i.e., axillary or retinal veins, <i>p</i> &lt; 0.001) and the overall number of TE episodes, was also significantly higher in PV patients than in the ET cohort (<i>p</i> &lt; 0.001), as reported in Table 1 below.</p><p>ET patients diagnosed with HTN, were more likely to have CVA than those without (of 119 HTN-affected, 15 had CVA, and of 441 HTN-negative, 28 had CVA, <i>p</i> = 0.032, Figure 1B, i) Patients with HTN were also more likely to have a venous thrombotic episode (of 119 HTN-affected, 21 had VTE, and of 441 HTN-negative, 44 had VTE, <i>p</i> = 0.021, Figure 1B, ii). Similarly, among PV patients who experienced a CVA, 39% (20) had HTN, demonstrating that, as would be expected, hypertension predisposes to CVA (20:63 vs. 31:246, <i>p</i> = 0.004, Figure 1B, iii). However, unlike ET, no correlation was reported between HTN and VTE (19:64 vs. 65:212, <i>p</i> &gt; 0.05, Figure 1B, iv).</p><p>Considering overall TE, for ET, 31.9% (38) of patients were affected by HTN, 22.2% (12) by HC, 19.2% (5) by DM, and 24.3% (9) had smoking habit reported a TE in their clinical history. Multivariate analysis of the ET cohort (considering HTN, HC, DM, and smoking habit) confirms the central role of HTN in increasing the risk of TE (OR: 2.5; 95% CI: 1.5–4.2; <i>p</i> &lt; 0.001). Regarding the PV cohort, 126 (35%) patients had a TE, and 44.6% (37) of HTN-affected patients experienced thrombosis. 48.1% (13) of HC, 35% (7) of DM, and 25% (5) of smoking habit patients experienced a TE. Applying a multivariate analysis (with the same cardiovascular risk factors described for ET), patients affected by HTN in the PV cohort show a higher risk of experiencing TE (OR: 1.5; 95% CI: 1.1–2.8; <i>p</i> &lt; 0.016).</p><p>To our knowledge, this is the first time a machine learning approach has been utilized to process and analyse large volume data in the ET and PV patient population, to provide valuable clinical insights. We describe a novel approach to cardiovascular risk assessment in patients with ET and PV, by using our NLP approach, we were able to analyse over 23,000 hematology documents that would have otherwise taken significant human time and labor. The advantage of our approach is that it can give near real-time updates of clinical events, which can inform patient management and risk prediction. Our manual validation demonstrated adequate performance in identifying a range of cardiovascular comorbidities and TE using this approach.</p><p>A previous report of 891 ET patients showed a prevalence of 5% for CVA, 2% for MI, and 4% for VTE, suggesting that our approach's detection rate is within acceptable limits.<span><sup>9</sup></span> Moreover, in an extensive study of Italian hematology centers, 235 patients with PV and 259 with ET were retrospectively evaluated, with a reported occurrence of thrombosis in 20.4% and 13%, respectively. Concerning CV risk factors, smoking was reported in 14.3%, HTN in 46.5%, HC in 12.5%, and DM in 8.2%, with results from these patient cohorts again consistent with our findings.<span><sup>10</sup></span></p><p>In keeping with previous studies our data suggest a higher risk for thrombotic events in PV, specifically regarding CVA, PVT, and VTE, when compared with the ET cohort. There was also a greater occurrence of overall thrombosis in the PV patient group. This is likely to partially reflect the greater frequency and higher variant allele frequency (VAF) of JAK2 mutation observed in PV patients. These data provide a basis for further mechanistic analysis to better define differences in thrombotic risk and assessment of the impact of driver mutations within disease groups is also warranted.</p><p>We provide a rare “real-world” report on the prevalence of comorbidities in this patient group. We have shown that hypertension is a comorbidity of great significance with regards to impact on thrombotic risk in MPN, with significantly increased risk of CVA in both PV and ET cohorts. This suggests a particular focus on controlling HTN is warranted and also highlights the need for health education for MPN patients to prevent the onset of hypertension. However, we were unable to assess differences between well-controlled and refractory or untreated hypertension using this approach. Our data support the routine assessment of cardiovascular comorbidities in defining the thrombotic risk in MPN patients, through the use of novel scoring systems such as QRISK-3. Studies should explore incorporating these comorbidities into predictive models to stratify patients based on thrombotic risk, as well as assess impact of targeted therapies on reducing associated risk.<span><sup>11</sup></span></p><p>A limitation of the NLP approach is that we have prespecified variables of interest for the model to detect, which prevents assessment of the impact of other associated factors, without development of an increasingly complex model. For example, it was not possible to assess the impact of driver mutation status on thrombotic risk using this approach in the current analysis.</p><p>Medical language is often complex, and shorthand or abbreviations used may be both department and region-dependent. For example a doctor documenting “AF” for atrial fibrillation could also mean “artificial feed” when used by dieticians. Using clinician-annotated documents, we are able to fine-tune the model to account for these variations. However, the performance of this approach is largely dependent on the quantity and the quality of training data and clinician annotations. Where there is large variation in the spelling, typing errors, context, or syntax, this will usually require more detailed annotations to perform well. This phenomenon is reflected in the somewhat moderate performance of “Thrombosis, not otherwise specified”. We used a broader SNOMED concept to capture thromboses that did not fit into our prespecified feature selection, but for this concept alone there were over 20 word/phrase annotation variations. The model performance could be improved to learn and generalize in broad concepts, but it will require further training over a larger number and range of documents.</p><p>Another limitation to the NLP approach is that it fails to capture conditions or statuses that are fluctuant or temporal, or those that are not explicitly stated. For example, for the concept of “Smoker,” MedCAT achieved a document-level F1 score of 0.87 (see Supporting Information S1: Table 1), however only achieved an F1 score of 0.64 and 0.66 in the ET and PV validation cohorts, respectively. This is because in a series of visits, a patient may well have stopped or restarted smoking, and smoking status may not be explicitly stated, for example, we do not always document that the patient is a “smoker,” instead a clinician may write “the patient smokes 20 cigarettes a day,” or “the patient only smokes socially.”</p><p>Finally, like all electronic health record research, the quality of NLP outputs is limited by the data quality of documented free text. For example, there was only one clinician annotation of “obesity” (see Supporting Information S1: Table 1), this would explain the lack of any cases of obesity being detected and suggests this co-morbidity may not be frequently documented, or alternatively this information may be coded in structured data (i.e., recorded weight and height measurements).</p><p>In addition, our study could be limited by the inclusion of analysis of patients with only one visit to our clinic. By analyzing these cases separately and validating them manually, we confirmed that the specificity and sensitivity of the natural processing language model are acceptable for ET and PV. Further studies are required to analyze the use of our NLP pipeline prospectively to define cardiovascular risk in MPN patients.</p><p>All authors have contributed substantially to the following. Andrea Duminuco, Joshua Au Yeung, Claire Harrison, and Patrick Harrington interpreted and analyzed the data and drafted the article: Raj Vaghela, Sukhraj Virdee, Claire Woodley, Susan Asirvatham, Natalia Curto-Garcia, Priya Sriskandarajah, Jennifer O'Sullivan, Hugues de Lavallade, Deepti Radia, Shahram Kordasti, and Giuseppe Palumbo revised the manuscript for intellectual content: All authors contributed to the article and approved the submitted version.</p><p>Hugues de Lavallade has received research grants and honoraria from Incyte and honoraria from Novartis and Pfizer. Claire Harrison has received speaker fees from Novartis, Jannsen, CTI, Celgene, Medscape and has served on the Advisory Board for Incyte, CTI, Sierra Oncology, Novartis, Celgene, Roche, AOP pharma, Geron and Astra Zenica, and is an Editor of HemaSphere. Patrick Harrington has received research funding from GSK, BMS, Novartis, Incyte, AOP, and Constellation and honoraria from GSK, Incyte, and Novartis.</p><p>This research received no funding.</p>","PeriodicalId":12982,"journal":{"name":"HemaSphere","volume":"8 8","pages":""},"PeriodicalIF":7.6000,"publicationDate":"2024-08-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11310405/pdf/","citationCount":"0","resultStr":"{\"title\":\"Development of a natural language processing pipeline for assessment of cardiovascular risk in myeloproliferative neoplasms\",\"authors\":\"Andrea Duminuco,&nbsp;Joshua Au Yeung,&nbsp;Raj Vaghela,&nbsp;Sukhraj Virdee,&nbsp;Claire Woodley,&nbsp;Susan Asirvatham,&nbsp;Natalia Curto-Garcia,&nbsp;Priya Sriskandarajah,&nbsp;Jennifer O'Sullivan,&nbsp;Hugues de Lavallade,&nbsp;Deepti Radia,&nbsp;Shahram Kordasti,&nbsp;Giuseppe Palumbo,&nbsp;Claire Harrison,&nbsp;Patrick Harrington\",\"doi\":\"10.1002/hem3.143\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<p>A central feature of myeloproliferative neoplasms (MPN) is an increased risk of cardiovascular thrombotic complications, and this is the primary determinant for the introduction of cytoreductive therapy.<span><sup>1</sup></span> The landmark ECLAP study in polycythemia vera (PV) patients, showed cardiovascular mortality accounted for 45% of all deaths, with a thrombosis incidence rate of 1.7/100 person/year and a cumulative incidence of 4.5% over a median follow-up of 2.8 years.<span><sup>2</sup></span></p><p>Natural language processing (NLP) is a branch of machine learning involving computational interpretation and analysis of human language. CogStack (https://github.com/CogStack), is an open-source software ecosystem, that retrieves structured and unstructured components of electronic health records (EHR). The Medical Concept Annotation Toolkit (MedCAT), the NLP component of CogStack, structures clinical free text by disambiguating and capturing synonyms, acronyms, and contextual details, such as negation, subject, and grammatical tense, and mapping text to medical Systematized Nomenclature of Medicine–Clinical Terms (SNOMED-CT) concepts. This technique is known as “named entity recognition and linkage” (NER+L). MedCAT has previously been used and validated in many studies to structure EHR data across a range of medical specialties for auditing, observational studies, de-identifying patient records, operational insights, disease modeling, and prediction.<span><sup>3-8</sup></span></p><p>We employed our NLP pipeline, Cogstack, and MedCAT, to determine the prevalence and impact of cardiovascular risk factors upon thrombotic events during follow-up. We used Cogstack to retrieve outpatient hematology clinic letters and hematology discharge letters. MedCAT was then used for NER+L of relevant clinical free-text to respective SNOMED-CT codes that were determined by two hematology specialists. The base MedCAT model was trained unsupervised on &gt;18 million EHR documents, and this was further fine-tuned using a 80:20 train:test split with 600 clinician-annotated MPN-specific documents. Total SNOMED-CT code counts were aggregated and grouped by individual patient, a unique threshold count was then applied to “infer” presence of the respective SNOMED code. In this process, hematology specialists read through clinical documents and manually highlight correct words or phrases detected by MedCAT that correspond to the SNOMED concept of interest.</p><p>We deploy a two-step validation process that has been well described.<span><sup>3, 8</sup></span> The first is to evaluate and validate the NER model performance on a document level demonstrating how accurately MedCAT is able to identify the medical concepts of interest. This involves hematology specialists annotating medical concepts and comparing this to the model NER outputs (Supporting Information S1: Table 1). The second step involves manual validation by creating a gold-standard real-world dataset. Two hematology specialists were randomly assigned to review a subsample of the patient sample's clinical notes (<i>n</i> = 112 [20%] in ET cohort, and <i>n</i> = 60 [17%] in the PV cohort) and, taking into account the entire clinical history, state whether the patient had presence or absence of the selected SNOMED concepts (Supporting Information S1: Tables 2 and 3). Finally, using the manual-validated data set, a threshold optimizer was used to find optimal concept counts for real-world F1 inference. The manual validation is crucial because whilst ML models can mislabel concepts on a document level, the key factor is how well a model performs in the real-world at recognizing the presence or absence of a clinical entity on a patient level.</p><p>The selected SNOMED concepts were cardiovascular risk factors, including hypertension (HTN), hypercholesterolemia (HC), diabetes mellitus (DM), smoker status, and obesity. We also assessed for cardiovascular events, including portal vein thrombosis (PVT), deep vein thrombosis (DVT), pulmonary embolism (PE), myocardial infarction (MI), stroke/cerebrovascular accident (CVA), cerebral sinus thrombosis (CST), and thrombosis not otherwise specified (NOS). A schematic representation of the process is represented in Figure 1A.</p><p>Data from 360 PV and 560 ET patients, reviewed at Guys' and St Thomas NHS Foundation Trust (GSTT) for at least one visit between January 2005 and April 2023, were evaluated (Supporting Information S1: Table 4). A total of 12905 documents from 560 ET patients (median 20 per patient, interquartile range [IQR], 8–34), and 11250 from 360 PV patients, (median 27 per patient, IQR, 11–47), were reviewed. In the manual validation data set (<i>n</i> = 112 patients for ET and <i>n</i> = 60 for PV), MedCAT achieved excellent real-world F1 scores with regards to specificity and sensitivity (Supporting Information S1: Tables 2 and 3).</p><p>For the ET cohort, using a threshold of &gt;2 mentions to define a positive population, HTN was identified in 21.3% (119) of patients, DM in 4.6% (26), MI in 3.6% (20), CVA in 7.7% (43), NOS thrombosis in 8% (45), DVT in 1.4% (8), PE in 1.8% (10), PVT in 1.3% (7) and positive smoking status in 6.6% (37) (Supporting Information S1: Figure 1A,B). HC was identified in 9.6% (54) using a threshold &gt;1. 52% (56) of patients with HC and 69.2% (18) of those with DM also had HTN. Obesity was not identified in any patients using this approach. Considering overall venous thromboembolic (VTE) occurrence, 11.6% (65) of patients reported an event. Of patients with CVA/MI, 58.1% (43)/55% (11) had this event pre- or at diagnosis, and 30.2% (13)/10% (2) while receiving cytoreductive therapy.</p><p>For the PV cohort, using a threshold of &gt;2 mentions to define the presence of the condition, HTN was identified in 23.1% (83) of cases, DM in 5.6% (20), MI in 3.1% (11), NOS thrombosis in 19.4% (70), DVT in 2.8% (10), PE in 2.8% (10), and PVT in 5% (18) with only one case of cerebral venous sinus thrombosis identified (0.3%) (Supporting Information S1: Figure 2A,B). Positive smoking habit was reported in 5.6% (20) of patients. Overall VTE was reported in 23.3% (84) of cases. Using a threshold of 1 mention, HC and CVA were detected in 7.5% (27) and 14.2% (51) patients, respectively. In keeping with the ET cohort, obesity was not identified in any patients. CVA/MI events occurred in 58.8% (30)/54.5 (6) cases prior to or at diagnosis and 23.5% (12)/9.1 (1) during cytoreductive treatment. Patients affected by DM suffered from HTN in 75% of cases (15), 66.7% (18) of HC patients had concomitant diagnosis of HTN, while 1.7% of the total had HTN, HC, and DM together.</p><p>TE were significantly higher in PV patients in the case of CVA (<i>p</i> = 0.002), PVT (<i>p</i> &lt; 0.001), venous thromboembolism (VTE, referring to any form of venous thrombotic event in an anatomical region that is not otherwise mentioned, i.e., axillary or retinal veins, <i>p</i> &lt; 0.001) and the overall number of TE episodes, was also significantly higher in PV patients than in the ET cohort (<i>p</i> &lt; 0.001), as reported in Table 1 below.</p><p>ET patients diagnosed with HTN, were more likely to have CVA than those without (of 119 HTN-affected, 15 had CVA, and of 441 HTN-negative, 28 had CVA, <i>p</i> = 0.032, Figure 1B, i) Patients with HTN were also more likely to have a venous thrombotic episode (of 119 HTN-affected, 21 had VTE, and of 441 HTN-negative, 44 had VTE, <i>p</i> = 0.021, Figure 1B, ii). Similarly, among PV patients who experienced a CVA, 39% (20) had HTN, demonstrating that, as would be expected, hypertension predisposes to CVA (20:63 vs. 31:246, <i>p</i> = 0.004, Figure 1B, iii). However, unlike ET, no correlation was reported between HTN and VTE (19:64 vs. 65:212, <i>p</i> &gt; 0.05, Figure 1B, iv).</p><p>Considering overall TE, for ET, 31.9% (38) of patients were affected by HTN, 22.2% (12) by HC, 19.2% (5) by DM, and 24.3% (9) had smoking habit reported a TE in their clinical history. Multivariate analysis of the ET cohort (considering HTN, HC, DM, and smoking habit) confirms the central role of HTN in increasing the risk of TE (OR: 2.5; 95% CI: 1.5–4.2; <i>p</i> &lt; 0.001). Regarding the PV cohort, 126 (35%) patients had a TE, and 44.6% (37) of HTN-affected patients experienced thrombosis. 48.1% (13) of HC, 35% (7) of DM, and 25% (5) of smoking habit patients experienced a TE. Applying a multivariate analysis (with the same cardiovascular risk factors described for ET), patients affected by HTN in the PV cohort show a higher risk of experiencing TE (OR: 1.5; 95% CI: 1.1–2.8; <i>p</i> &lt; 0.016).</p><p>To our knowledge, this is the first time a machine learning approach has been utilized to process and analyse large volume data in the ET and PV patient population, to provide valuable clinical insights. We describe a novel approach to cardiovascular risk assessment in patients with ET and PV, by using our NLP approach, we were able to analyse over 23,000 hematology documents that would have otherwise taken significant human time and labor. The advantage of our approach is that it can give near real-time updates of clinical events, which can inform patient management and risk prediction. Our manual validation demonstrated adequate performance in identifying a range of cardiovascular comorbidities and TE using this approach.</p><p>A previous report of 891 ET patients showed a prevalence of 5% for CVA, 2% for MI, and 4% for VTE, suggesting that our approach's detection rate is within acceptable limits.<span><sup>9</sup></span> Moreover, in an extensive study of Italian hematology centers, 235 patients with PV and 259 with ET were retrospectively evaluated, with a reported occurrence of thrombosis in 20.4% and 13%, respectively. Concerning CV risk factors, smoking was reported in 14.3%, HTN in 46.5%, HC in 12.5%, and DM in 8.2%, with results from these patient cohorts again consistent with our findings.<span><sup>10</sup></span></p><p>In keeping with previous studies our data suggest a higher risk for thrombotic events in PV, specifically regarding CVA, PVT, and VTE, when compared with the ET cohort. There was also a greater occurrence of overall thrombosis in the PV patient group. This is likely to partially reflect the greater frequency and higher variant allele frequency (VAF) of JAK2 mutation observed in PV patients. These data provide a basis for further mechanistic analysis to better define differences in thrombotic risk and assessment of the impact of driver mutations within disease groups is also warranted.</p><p>We provide a rare “real-world” report on the prevalence of comorbidities in this patient group. We have shown that hypertension is a comorbidity of great significance with regards to impact on thrombotic risk in MPN, with significantly increased risk of CVA in both PV and ET cohorts. This suggests a particular focus on controlling HTN is warranted and also highlights the need for health education for MPN patients to prevent the onset of hypertension. However, we were unable to assess differences between well-controlled and refractory or untreated hypertension using this approach. Our data support the routine assessment of cardiovascular comorbidities in defining the thrombotic risk in MPN patients, through the use of novel scoring systems such as QRISK-3. Studies should explore incorporating these comorbidities into predictive models to stratify patients based on thrombotic risk, as well as assess impact of targeted therapies on reducing associated risk.<span><sup>11</sup></span></p><p>A limitation of the NLP approach is that we have prespecified variables of interest for the model to detect, which prevents assessment of the impact of other associated factors, without development of an increasingly complex model. For example, it was not possible to assess the impact of driver mutation status on thrombotic risk using this approach in the current analysis.</p><p>Medical language is often complex, and shorthand or abbreviations used may be both department and region-dependent. For example a doctor documenting “AF” for atrial fibrillation could also mean “artificial feed” when used by dieticians. Using clinician-annotated documents, we are able to fine-tune the model to account for these variations. However, the performance of this approach is largely dependent on the quantity and the quality of training data and clinician annotations. Where there is large variation in the spelling, typing errors, context, or syntax, this will usually require more detailed annotations to perform well. This phenomenon is reflected in the somewhat moderate performance of “Thrombosis, not otherwise specified”. We used a broader SNOMED concept to capture thromboses that did not fit into our prespecified feature selection, but for this concept alone there were over 20 word/phrase annotation variations. The model performance could be improved to learn and generalize in broad concepts, but it will require further training over a larger number and range of documents.</p><p>Another limitation to the NLP approach is that it fails to capture conditions or statuses that are fluctuant or temporal, or those that are not explicitly stated. For example, for the concept of “Smoker,” MedCAT achieved a document-level F1 score of 0.87 (see Supporting Information S1: Table 1), however only achieved an F1 score of 0.64 and 0.66 in the ET and PV validation cohorts, respectively. This is because in a series of visits, a patient may well have stopped or restarted smoking, and smoking status may not be explicitly stated, for example, we do not always document that the patient is a “smoker,” instead a clinician may write “the patient smokes 20 cigarettes a day,” or “the patient only smokes socially.”</p><p>Finally, like all electronic health record research, the quality of NLP outputs is limited by the data quality of documented free text. For example, there was only one clinician annotation of “obesity” (see Supporting Information S1: Table 1), this would explain the lack of any cases of obesity being detected and suggests this co-morbidity may not be frequently documented, or alternatively this information may be coded in structured data (i.e., recorded weight and height measurements).</p><p>In addition, our study could be limited by the inclusion of analysis of patients with only one visit to our clinic. By analyzing these cases separately and validating them manually, we confirmed that the specificity and sensitivity of the natural processing language model are acceptable for ET and PV. Further studies are required to analyze the use of our NLP pipeline prospectively to define cardiovascular risk in MPN patients.</p><p>All authors have contributed substantially to the following. Andrea Duminuco, Joshua Au Yeung, Claire Harrison, and Patrick Harrington interpreted and analyzed the data and drafted the article: Raj Vaghela, Sukhraj Virdee, Claire Woodley, Susan Asirvatham, Natalia Curto-Garcia, Priya Sriskandarajah, Jennifer O'Sullivan, Hugues de Lavallade, Deepti Radia, Shahram Kordasti, and Giuseppe Palumbo revised the manuscript for intellectual content: All authors contributed to the article and approved the submitted version.</p><p>Hugues de Lavallade has received research grants and honoraria from Incyte and honoraria from Novartis and Pfizer. Claire Harrison has received speaker fees from Novartis, Jannsen, CTI, Celgene, Medscape and has served on the Advisory Board for Incyte, CTI, Sierra Oncology, Novartis, Celgene, Roche, AOP pharma, Geron and Astra Zenica, and is an Editor of HemaSphere. Patrick Harrington has received research funding from GSK, BMS, Novartis, Incyte, AOP, and Constellation and honoraria from GSK, Incyte, and Novartis.</p><p>This research received no funding.</p>\",\"PeriodicalId\":12982,\"journal\":{\"name\":\"HemaSphere\",\"volume\":\"8 8\",\"pages\":\"\"},\"PeriodicalIF\":7.6000,\"publicationDate\":\"2024-08-08\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11310405/pdf/\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"HemaSphere\",\"FirstCategoryId\":\"3\",\"ListUrlMain\":\"https://onlinelibrary.wiley.com/doi/10.1002/hem3.143\",\"RegionNum\":2,\"RegionCategory\":\"医学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"HEMATOLOGY\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"HemaSphere","FirstCategoryId":"3","ListUrlMain":"https://onlinelibrary.wiley.com/doi/10.1002/hem3.143","RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"HEMATOLOGY","Score":null,"Total":0}
引用次数: 0

摘要

骨髓增生性肿瘤(MPN)的一个主要特征是心血管血栓并发症风险增加,这也是采用细胞再生疗法的主要决定因素。对多发性红细胞增多症(PV)患者进行的具有里程碑意义的 ECLAP 研究显示,心血管死亡占所有死亡的 45%,血栓形成发生率为 1.7/100人/年,中位随访 2.8 年的累积发生率为 4.5%。CogStack (https://github.com/CogStack) 是一个开源软件生态系统,可检索电子健康记录 (EHR) 中的结构化和非结构化组件。医学概念注释工具包(MedCAT)是 CogStack 的 NLP 组件,它通过消歧和捕捉同义词、缩略词和上下文细节(如否定词、主语和语法时态),将文本映射到医学术语系统化命名法(SNOMED-CT)概念,从而构建临床自由文本。这种技术被称为 "命名实体识别和链接"(NER+L)。MedCAT 此前已在多项研究中得到应用和验证,可用于构建一系列医疗专科的电子病历数据,以进行审计、观察研究、去身份化患者记录、业务洞察、疾病建模和预测3-8。我们使用 Cogstack 来检索血液科门诊信件和血液科出院信件。然后使用 MedCAT 对相关的临床自由文本进行 NER+L 处理,使之与两位血液学专家确定的 SNOMED-CT 代码相对应。基础 MedCAT 模型在 1800 万份电子病历文档上进行了无监督训练,并通过 600 份临床医生注释的 MPN 特定文档,以 80:20 的训练与测试比例对其进行了进一步微调。对 SNOMED-CT 代码的总计数进行汇总,并按单个患者进行分组,然后应用独特的阈值计数来 "推断 "是否存在相应的 SNOMED 代码。在这一过程中,血液学专家通读临床文档,并手动高亮显示 MedCAT 检测到的与相关 SNOMED 概念相对应的正确单词或短语。我们采用了两步验证流程,这一流程已得到充分描述。这需要血液学专家对医学概念进行注释,并将其与 NER 模型的输出结果进行比较(佐证资料 S1:表 1)。第二步是通过创建黄金标准真实世界数据集进行人工验证。随机指派两名血液学专家审阅患者样本临床笔记的子样本(ET 队列中的样本数为 112 [20%],PV 队列中的样本数为 60 [17%]),并结合整个临床病史,说明患者是否存在所选的 SNOMED 概念(佐证资料 S1:表 2 和表 3)。最后,利用人工验证的数据集,使用阈值优化器为真实世界的 F1 推断找到最佳概念数。人工验证至关重要,因为虽然 ML 模型可能会在文档层面上错误标注概念,但关键因素是模型在真实世界中识别患者层面上临床实体存在与否的性能如何。所选的 SNOMED 概念是心血管风险因素,包括高血压 (HTN)、高胆固醇血症 (HC)、糖尿病 (DM)、吸烟状态和肥胖。我们还对心血管事件进行了评估,包括门静脉血栓 (PVT)、深静脉血栓 (DVT)、肺栓塞 (PE)、心肌梗塞 (MI)、中风/脑血管意外 (CVA)、脑窦血栓 (CST) 和未注明的血栓 (NOS)。对 2005 年 1 月至 2023 年 4 月期间在 Guys' and St Thomas NHS Foundation Trust (GSTT) 至少就诊过一次的 360 名 PV 和 560 名 ET 患者的数据进行了评估(佐证资料 S1:表 4)。共审查了 560 名 ET 患者的 12905 份文件(中位数为每名患者 20 份,四分位数间距 [IQR],8-34)和 360 名 PV 患者的 11250 份文件(中位数为每名患者 27 份,四分位数间距 [IQR],11-47)。在人工验证数据集(ET 患者人数为 112 人,PV 患者人数为 60 人)中,MedCAT 在特异性和灵敏度方面取得了极佳的真实世界 F1 分数(佐证资料 S1:表 2 和表 3)。在 ET 队列中,使用阈值&gt;2 次提及来定义阳性人群,21.3%(119 例)的患者发现了高血压,4.6%(26 例)的患者发现了糖尿病,3.6%(20 例)的患者发现了心肌梗塞,7.7%(43 例)的患者发现了 CVA,8%(45 例)的患者发现了 NOS 血栓,1.4%(8 例)的患者发现了深静脉血栓,1.4%(8 例)的患者发现了 PE。 Hugues de Lavallade 曾获得 Incyte 的研究基金和酬金,以及诺华和辉瑞的酬金。克莱尔-哈里森(Claire Harrison)从诺华、杨森、CTI、Celgene、Medscape 等公司获得演讲酬金,并担任过 Incyte、CTI、Sierra Oncology、诺华、Celgene、罗氏、AOP pharma、Geron 和 Astra Zenica 的顾问委员会成员,还是 HemaSphere 的编辑。Patrick Harrington 从葛兰素史克、BMS、诺华、Incyte、AOP 和 Constellation 获得研究经费,并从葛兰素史克、Incyte 和诺华获得酬金。
本文章由计算机程序翻译,如有差异,请以英文原文为准。

Development of a natural language processing pipeline for assessment of cardiovascular risk in myeloproliferative neoplasms

Development of a natural language processing pipeline for assessment of cardiovascular risk in myeloproliferative neoplasms

A central feature of myeloproliferative neoplasms (MPN) is an increased risk of cardiovascular thrombotic complications, and this is the primary determinant for the introduction of cytoreductive therapy.1 The landmark ECLAP study in polycythemia vera (PV) patients, showed cardiovascular mortality accounted for 45% of all deaths, with a thrombosis incidence rate of 1.7/100 person/year and a cumulative incidence of 4.5% over a median follow-up of 2.8 years.2

Natural language processing (NLP) is a branch of machine learning involving computational interpretation and analysis of human language. CogStack (https://github.com/CogStack), is an open-source software ecosystem, that retrieves structured and unstructured components of electronic health records (EHR). The Medical Concept Annotation Toolkit (MedCAT), the NLP component of CogStack, structures clinical free text by disambiguating and capturing synonyms, acronyms, and contextual details, such as negation, subject, and grammatical tense, and mapping text to medical Systematized Nomenclature of Medicine–Clinical Terms (SNOMED-CT) concepts. This technique is known as “named entity recognition and linkage” (NER+L). MedCAT has previously been used and validated in many studies to structure EHR data across a range of medical specialties for auditing, observational studies, de-identifying patient records, operational insights, disease modeling, and prediction.3-8

We employed our NLP pipeline, Cogstack, and MedCAT, to determine the prevalence and impact of cardiovascular risk factors upon thrombotic events during follow-up. We used Cogstack to retrieve outpatient hematology clinic letters and hematology discharge letters. MedCAT was then used for NER+L of relevant clinical free-text to respective SNOMED-CT codes that were determined by two hematology specialists. The base MedCAT model was trained unsupervised on >18 million EHR documents, and this was further fine-tuned using a 80:20 train:test split with 600 clinician-annotated MPN-specific documents. Total SNOMED-CT code counts were aggregated and grouped by individual patient, a unique threshold count was then applied to “infer” presence of the respective SNOMED code. In this process, hematology specialists read through clinical documents and manually highlight correct words or phrases detected by MedCAT that correspond to the SNOMED concept of interest.

We deploy a two-step validation process that has been well described.3, 8 The first is to evaluate and validate the NER model performance on a document level demonstrating how accurately MedCAT is able to identify the medical concepts of interest. This involves hematology specialists annotating medical concepts and comparing this to the model NER outputs (Supporting Information S1: Table 1). The second step involves manual validation by creating a gold-standard real-world dataset. Two hematology specialists were randomly assigned to review a subsample of the patient sample's clinical notes (n = 112 [20%] in ET cohort, and n = 60 [17%] in the PV cohort) and, taking into account the entire clinical history, state whether the patient had presence or absence of the selected SNOMED concepts (Supporting Information S1: Tables 2 and 3). Finally, using the manual-validated data set, a threshold optimizer was used to find optimal concept counts for real-world F1 inference. The manual validation is crucial because whilst ML models can mislabel concepts on a document level, the key factor is how well a model performs in the real-world at recognizing the presence or absence of a clinical entity on a patient level.

The selected SNOMED concepts were cardiovascular risk factors, including hypertension (HTN), hypercholesterolemia (HC), diabetes mellitus (DM), smoker status, and obesity. We also assessed for cardiovascular events, including portal vein thrombosis (PVT), deep vein thrombosis (DVT), pulmonary embolism (PE), myocardial infarction (MI), stroke/cerebrovascular accident (CVA), cerebral sinus thrombosis (CST), and thrombosis not otherwise specified (NOS). A schematic representation of the process is represented in Figure 1A.

Data from 360 PV and 560 ET patients, reviewed at Guys' and St Thomas NHS Foundation Trust (GSTT) for at least one visit between January 2005 and April 2023, were evaluated (Supporting Information S1: Table 4). A total of 12905 documents from 560 ET patients (median 20 per patient, interquartile range [IQR], 8–34), and 11250 from 360 PV patients, (median 27 per patient, IQR, 11–47), were reviewed. In the manual validation data set (n = 112 patients for ET and n = 60 for PV), MedCAT achieved excellent real-world F1 scores with regards to specificity and sensitivity (Supporting Information S1: Tables 2 and 3).

For the ET cohort, using a threshold of >2 mentions to define a positive population, HTN was identified in 21.3% (119) of patients, DM in 4.6% (26), MI in 3.6% (20), CVA in 7.7% (43), NOS thrombosis in 8% (45), DVT in 1.4% (8), PE in 1.8% (10), PVT in 1.3% (7) and positive smoking status in 6.6% (37) (Supporting Information S1: Figure 1A,B). HC was identified in 9.6% (54) using a threshold >1. 52% (56) of patients with HC and 69.2% (18) of those with DM also had HTN. Obesity was not identified in any patients using this approach. Considering overall venous thromboembolic (VTE) occurrence, 11.6% (65) of patients reported an event. Of patients with CVA/MI, 58.1% (43)/55% (11) had this event pre- or at diagnosis, and 30.2% (13)/10% (2) while receiving cytoreductive therapy.

For the PV cohort, using a threshold of >2 mentions to define the presence of the condition, HTN was identified in 23.1% (83) of cases, DM in 5.6% (20), MI in 3.1% (11), NOS thrombosis in 19.4% (70), DVT in 2.8% (10), PE in 2.8% (10), and PVT in 5% (18) with only one case of cerebral venous sinus thrombosis identified (0.3%) (Supporting Information S1: Figure 2A,B). Positive smoking habit was reported in 5.6% (20) of patients. Overall VTE was reported in 23.3% (84) of cases. Using a threshold of 1 mention, HC and CVA were detected in 7.5% (27) and 14.2% (51) patients, respectively. In keeping with the ET cohort, obesity was not identified in any patients. CVA/MI events occurred in 58.8% (30)/54.5 (6) cases prior to or at diagnosis and 23.5% (12)/9.1 (1) during cytoreductive treatment. Patients affected by DM suffered from HTN in 75% of cases (15), 66.7% (18) of HC patients had concomitant diagnosis of HTN, while 1.7% of the total had HTN, HC, and DM together.

TE were significantly higher in PV patients in the case of CVA (p = 0.002), PVT (p < 0.001), venous thromboembolism (VTE, referring to any form of venous thrombotic event in an anatomical region that is not otherwise mentioned, i.e., axillary or retinal veins, p < 0.001) and the overall number of TE episodes, was also significantly higher in PV patients than in the ET cohort (p < 0.001), as reported in Table 1 below.

ET patients diagnosed with HTN, were more likely to have CVA than those without (of 119 HTN-affected, 15 had CVA, and of 441 HTN-negative, 28 had CVA, p = 0.032, Figure 1B, i) Patients with HTN were also more likely to have a venous thrombotic episode (of 119 HTN-affected, 21 had VTE, and of 441 HTN-negative, 44 had VTE, p = 0.021, Figure 1B, ii). Similarly, among PV patients who experienced a CVA, 39% (20) had HTN, demonstrating that, as would be expected, hypertension predisposes to CVA (20:63 vs. 31:246, p = 0.004, Figure 1B, iii). However, unlike ET, no correlation was reported between HTN and VTE (19:64 vs. 65:212, p > 0.05, Figure 1B, iv).

Considering overall TE, for ET, 31.9% (38) of patients were affected by HTN, 22.2% (12) by HC, 19.2% (5) by DM, and 24.3% (9) had smoking habit reported a TE in their clinical history. Multivariate analysis of the ET cohort (considering HTN, HC, DM, and smoking habit) confirms the central role of HTN in increasing the risk of TE (OR: 2.5; 95% CI: 1.5–4.2; p < 0.001). Regarding the PV cohort, 126 (35%) patients had a TE, and 44.6% (37) of HTN-affected patients experienced thrombosis. 48.1% (13) of HC, 35% (7) of DM, and 25% (5) of smoking habit patients experienced a TE. Applying a multivariate analysis (with the same cardiovascular risk factors described for ET), patients affected by HTN in the PV cohort show a higher risk of experiencing TE (OR: 1.5; 95% CI: 1.1–2.8; p < 0.016).

To our knowledge, this is the first time a machine learning approach has been utilized to process and analyse large volume data in the ET and PV patient population, to provide valuable clinical insights. We describe a novel approach to cardiovascular risk assessment in patients with ET and PV, by using our NLP approach, we were able to analyse over 23,000 hematology documents that would have otherwise taken significant human time and labor. The advantage of our approach is that it can give near real-time updates of clinical events, which can inform patient management and risk prediction. Our manual validation demonstrated adequate performance in identifying a range of cardiovascular comorbidities and TE using this approach.

A previous report of 891 ET patients showed a prevalence of 5% for CVA, 2% for MI, and 4% for VTE, suggesting that our approach's detection rate is within acceptable limits.9 Moreover, in an extensive study of Italian hematology centers, 235 patients with PV and 259 with ET were retrospectively evaluated, with a reported occurrence of thrombosis in 20.4% and 13%, respectively. Concerning CV risk factors, smoking was reported in 14.3%, HTN in 46.5%, HC in 12.5%, and DM in 8.2%, with results from these patient cohorts again consistent with our findings.10

In keeping with previous studies our data suggest a higher risk for thrombotic events in PV, specifically regarding CVA, PVT, and VTE, when compared with the ET cohort. There was also a greater occurrence of overall thrombosis in the PV patient group. This is likely to partially reflect the greater frequency and higher variant allele frequency (VAF) of JAK2 mutation observed in PV patients. These data provide a basis for further mechanistic analysis to better define differences in thrombotic risk and assessment of the impact of driver mutations within disease groups is also warranted.

We provide a rare “real-world” report on the prevalence of comorbidities in this patient group. We have shown that hypertension is a comorbidity of great significance with regards to impact on thrombotic risk in MPN, with significantly increased risk of CVA in both PV and ET cohorts. This suggests a particular focus on controlling HTN is warranted and also highlights the need for health education for MPN patients to prevent the onset of hypertension. However, we were unable to assess differences between well-controlled and refractory or untreated hypertension using this approach. Our data support the routine assessment of cardiovascular comorbidities in defining the thrombotic risk in MPN patients, through the use of novel scoring systems such as QRISK-3. Studies should explore incorporating these comorbidities into predictive models to stratify patients based on thrombotic risk, as well as assess impact of targeted therapies on reducing associated risk.11

A limitation of the NLP approach is that we have prespecified variables of interest for the model to detect, which prevents assessment of the impact of other associated factors, without development of an increasingly complex model. For example, it was not possible to assess the impact of driver mutation status on thrombotic risk using this approach in the current analysis.

Medical language is often complex, and shorthand or abbreviations used may be both department and region-dependent. For example a doctor documenting “AF” for atrial fibrillation could also mean “artificial feed” when used by dieticians. Using clinician-annotated documents, we are able to fine-tune the model to account for these variations. However, the performance of this approach is largely dependent on the quantity and the quality of training data and clinician annotations. Where there is large variation in the spelling, typing errors, context, or syntax, this will usually require more detailed annotations to perform well. This phenomenon is reflected in the somewhat moderate performance of “Thrombosis, not otherwise specified”. We used a broader SNOMED concept to capture thromboses that did not fit into our prespecified feature selection, but for this concept alone there were over 20 word/phrase annotation variations. The model performance could be improved to learn and generalize in broad concepts, but it will require further training over a larger number and range of documents.

Another limitation to the NLP approach is that it fails to capture conditions or statuses that are fluctuant or temporal, or those that are not explicitly stated. For example, for the concept of “Smoker,” MedCAT achieved a document-level F1 score of 0.87 (see Supporting Information S1: Table 1), however only achieved an F1 score of 0.64 and 0.66 in the ET and PV validation cohorts, respectively. This is because in a series of visits, a patient may well have stopped or restarted smoking, and smoking status may not be explicitly stated, for example, we do not always document that the patient is a “smoker,” instead a clinician may write “the patient smokes 20 cigarettes a day,” or “the patient only smokes socially.”

Finally, like all electronic health record research, the quality of NLP outputs is limited by the data quality of documented free text. For example, there was only one clinician annotation of “obesity” (see Supporting Information S1: Table 1), this would explain the lack of any cases of obesity being detected and suggests this co-morbidity may not be frequently documented, or alternatively this information may be coded in structured data (i.e., recorded weight and height measurements).

In addition, our study could be limited by the inclusion of analysis of patients with only one visit to our clinic. By analyzing these cases separately and validating them manually, we confirmed that the specificity and sensitivity of the natural processing language model are acceptable for ET and PV. Further studies are required to analyze the use of our NLP pipeline prospectively to define cardiovascular risk in MPN patients.

All authors have contributed substantially to the following. Andrea Duminuco, Joshua Au Yeung, Claire Harrison, and Patrick Harrington interpreted and analyzed the data and drafted the article: Raj Vaghela, Sukhraj Virdee, Claire Woodley, Susan Asirvatham, Natalia Curto-Garcia, Priya Sriskandarajah, Jennifer O'Sullivan, Hugues de Lavallade, Deepti Radia, Shahram Kordasti, and Giuseppe Palumbo revised the manuscript for intellectual content: All authors contributed to the article and approved the submitted version.

Hugues de Lavallade has received research grants and honoraria from Incyte and honoraria from Novartis and Pfizer. Claire Harrison has received speaker fees from Novartis, Jannsen, CTI, Celgene, Medscape and has served on the Advisory Board for Incyte, CTI, Sierra Oncology, Novartis, Celgene, Roche, AOP pharma, Geron and Astra Zenica, and is an Editor of HemaSphere. Patrick Harrington has received research funding from GSK, BMS, Novartis, Incyte, AOP, and Constellation and honoraria from GSK, Incyte, and Novartis.

This research received no funding.

求助全文
通过发布文献求助,成功后即可免费获取论文全文。 去求助
来源期刊
HemaSphere
HemaSphere Medicine-Hematology
CiteScore
6.10
自引率
4.50%
发文量
2776
审稿时长
7 weeks
期刊介绍: HemaSphere, as a publication, is dedicated to disseminating the outcomes of profoundly pertinent basic, translational, and clinical research endeavors within the field of hematology. The journal actively seeks robust studies that unveil novel discoveries with significant ramifications for hematology. In addition to original research, HemaSphere features review articles and guideline articles that furnish lucid synopses and discussions of emerging developments, along with recommendations for patient care. Positioned as the foremost resource in hematology, HemaSphere augments its offerings with specialized sections like HemaTopics and HemaPolicy. These segments engender insightful dialogues covering a spectrum of hematology-related topics, including digestible summaries of pivotal articles, updates on new therapies, deliberations on European policy matters, and other noteworthy news items within the field. Steering the course of HemaSphere are Editor in Chief Jan Cools and Deputy Editor in Chief Claire Harrison, alongside the guidance of an esteemed Editorial Board comprising international luminaries in both research and clinical realms, each representing diverse areas of hematologic expertise.
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
确定
请完成安全验证×
copy
已复制链接
快去分享给好友吧!
我知道了
右上角分享
点击右上角分享
0
联系我们:info@booksci.cn Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。 Copyright © 2023 布克学术 All rights reserved.
京ICP备2023020795号-1
ghs 京公网安备 11010802042870号
Book学术文献互助
Book学术文献互助群
群 号:481959085
Book学术官方微信