Central nervous system manifestations in acute and chronic graft-versus-host disease.

Despite the growing evidence supporting the existence of CNS involvement in acute and chronic graft-versus-host disease (CNS-GvHD), the characteristics and course of the disease are still largely unknown. In this multicenter retrospective study, we analyzed the clinical, biological, radiological, and histopathological characteristics, as well as the clinical course of 66 patients diagnosed with possible CNS-GvHD (pCNS-GvHD), selected by predetermined diagnostic criteria. Results were then contrasted depending on whether pCNS-GvHD occurred before or after day 100 following allogeneic hematopoietic stem cell transplantation. Median time between hematopoietic stem cell transplantation and pCNS-GvHD onset was 149 days (IQ25-75 48-321), and pCNS-GvHD onset occurred before day 100 following transplantation in 44% of patients. The most frequent findings at presentation were cognitive impairment (41%), paresis (21%), altered consciousness (20%), sensory impairment (18%), and headache (15%). Clinical presentation did not significantly differ between patients with pCNS-GvHD occurring before or after day 100 following transplantation. Brain MRI found abnormalities compatible with the clinical picture in 57% of patients, while CT detected abnormalities in only 7%. Seven patients had documented spinal cord MRI abnormalities, all of them with pCNS-GvHD occurring after day 100 following transplantation. In the cerebrospinal fluid, white blood cell count was increased in 56% of the population (median 18 cells/µL). Histopathological analyses were performed on 12 specimens and were suggestive of pCNS-GvHD in 10. All compatible specimens showed parenchymal and perivascular infiltration by CD3+ and CD163+ cells. Immunosuppressive therapy was prescribed in 97% of patients, achieving complete clinical response in 27%, partial improvement in 47% and stable disease in 6%. Response to immunosuppressive therapy did not significantly differ between patients with pCNS-GvHD occurring before or after day 100 following transplantation. Clinical relapse was observed in 31% of patients who initially responded to treatment. One-year overall survival following pCNS-GvHD onset was 41%. Onset before day 100 following hematopoietic stem cell transplantation (HR [95%CI]: 2.1 [1.0-4.5]; P=0.041) and altered consciousness at initial presentation (HR [95%CI]: 3.0 [1.3-6.7]; P=0.0077) were associated with a reduced one-year overall survival probability. Among surviving patients, 61% had neurological sequelae. This study supports that immune-mediated CNS manifestations may occur following allo-HSCT. These can be associated with both acute and chronic GvHD and carry a grim prognosis. The clinical presentation as well as the radiological and biological findings appear variable.

Ibrutinib for therapy of steroid-refractory chronic graft vs. host disease: A multicenter real-world analysis.

To examine activity of ibrutinib in steroid-refractory chronic GVHD (SR-cGVHD) after FDA approval, we conducted a multicenter retrospective study. Data were standardly collected (N=270 from 19 centers). Involved organs included skin (75%), eye (61%), mouth (54%), joint/fascia (47%), GI (26%), lung (27%), liver (19%), genital (7%), other (4.4%). NIH severity was mild in 5.7%, moderate 42%, severe 53%. 39% had overlap subtype. KPS was ≥ 80% in 72%. Median prednisone (mg/kg) was 0.21 (0-2.27). Ibrutinib was started at median of 18.2 months after cGVHD onset and in earlier lines of therapy (2nd line: 26%, 3rd: 30%, 4th: 21%, 5th: 9.6%, 6th: 10%, 7th or higher: 1.2%)). Among evaluable subjects, the 6 month NIH overall response rate (CR/PR) was 45% (PR 42%, CR 3%). Median duration of response was 15 months (range 1-46). Liver involvement had association with 6 month ORR (multivariate (MVA) OR 5.49 (95% CI 2.3-14.2, p <0.001). Best overall response was 56%, with most (86%) achieving by 1-3 months. With median follow up for survivors of 30.5 months, FFS was 59% (53-65%) at 6 months and 41% (36-48%) at 12 months. On MVA, increased age (HR 1.01, 95% CI 1.0-1.02, p=0.033), higher baseline prednisone (HR 1.92, 1.09-3.38, p=0.032), and lung involvement (HR 1.58, 1.1-2.28, p=0.016) had worse FFS. Ibrutinib discontinuation was most commonly due to progressive cGVHD (44%) or toxicity (42%). These data support that ibrutinib has activity in SR-cGVHD, provide new insight into factors associated with response and FFS, and demonstrate the toxicity profile associated with discontinuation.

Access to Allogeneic Cell Transplantation Based on Donor Search Prognosis: An Interventional Trial.

Importance: Patients requiring allogeneic hematopoietic cell transplantation have variable likelihoods of identifying an 8/8 HLA-matched unrelated donor. A Search Prognosis calculator can estimate the likelihood. Objective: To determine if using a search algorithm based on donor search prognosis can result in similar incidence of transplant between patients Very Likely (>90%) vs Very Unlikely (<10%) to have a matched unrelated donor. Design: This interventional trial utilized a Search Prognosis-based biologic assignment algorithm to guide donor selection. Trial enrollment from June 13, 2019-May 13, 2022; analysis of data as of September 7, 2023 with median follow-up post-evaluability of 14.5 months. Settings: National multi-center Blood and Marrow Transplantation Clinical Trials Network 1702 study of US participating transplant centers. Participants: Acute myeloid and lymphoid leukemias, myelodysplastic syndrome, Hodgkin's and non-Hodgkin's lymphomas, severe aplastic anemia, and sickle cell disease patients referred to participating transplant centers were invited to participate. 2225 patients were enrolled and 1751 were declared evaluable for this study. Patients were declared evaluable once it was determined no suitable HLA-matched related donor was available. Intervention: Patients assigned to the Very Likely arm were to proceed with matched unrelated donor, while Very Unlikely were to utilize alternative donors. A third stratum, Less Likely (~25%) to find a matched unrelated donor, were observed under standard center practices, but were not part of the primary objective. Main Outcome: Cumulative incidence of transplantation by Search Prognosis arm. Results: Evaluable patients included 1751 of which 413 (24%) were from racial/ethnic minorities. Search prognosis was 958 (55%) Very Likely, 517 (30%) Less Likely and 276 (16%) Very Unlikely. 1171 (67%) received HCT, 384 (22%) died without HCT, and 196 (11%) remained alive without HCT. Among the 1,234 patients, the adjusted cumulative incidence (95% CI) of HCT at 6-months was 59.8% (56.7-62.8) in the Very Likely group versus 52.3% (46.1-58.5) in the Very Unlikely (P=0.113). Conclusions: A prospective Search Prognosis-based algorithm can be effectively implemented in a national multicenter clinical trial. This approach resulted in rapid alternative donor identification and comparable rates of HCT in patients Very Likely and Very Unlikely to find a matched unrelated donor.

Developing an automated algorithm for identification of children and adolescents with diabetes using electronic health records from the OneFlorida+ clinical research network.

AIM: To develop an automated computable phenotype (CP) algorithm for identifying diabetes cases in children and adolescents using electronic health records (EHRs) from the UF Health System. MATERIALS AND METHODS: The CP algorithm was iteratively derived based on structured data from EHRs (UF Health System 2012-2020). We randomly selected 536 presumed cases among individuals aged <18 years who had (1) glycated haemoglobin levels ≥ 6.5%; or (2) fasting glucose levels ≥126 mg/dL; or (3) random plasma glucose levels ≥200 mg/dL; or (4) a diabetes-related diagnosis code from an inpatient or outpatient encounter; or (5) prescribed, administered, or dispensed diabetes-related medication. Four reviewers independently reviewed the patient charts to determine diabetes status and type. RESULTS: Presumed cases without type 1 (T1D) or type 2 diabetes (T2D) diagnosis codes were categorized as non-diabetes/other types of diabetes. The rest were categorized as T1D if the most recent diagnosis was T1D, or otherwise categorized as T2D if the most recent diagnosis was T2D. Next, we applied a list of diagnoses and procedures that can determine diabetes type (e.g., steroid use suggests induced diabetes) to correct misclassifications from Step 1. Among the 536 reviewed cases, 159 and 64 had T1D and T2D, respectively. The sensitivity, specificity, and positive predictive values of the CP algorithm were 94%, 98% and 96%, respectively, for T1D and 95%, 95% and 73% for T2D. CONCLUSION: We developed a highly accurate EHR-based CP for diabetes in youth based on EHR data from UF Health. Consistent with prior studies, T2D was more difficult to identify using these methods.

Differences in Acute Graft-Versus-Host Disease (GVHD) Severity and Its Outcomes Between Black and White Patients.

Acute graft-versus-host disease (GVHD) is a significant complication following hematopoietic stem cell transplantation (HCT). Although recent advancements in GVHD prophylaxis have resulted in successful HCT across HLA barriers and expanded access to HCT for racial minorities, less is known about how race affects the severity and outcomes of acute GVHD. This study examines differences in the clinical course of acute GVHD and the prognostic value of GVHD biomarkers for Black and White recipients. We conducted a retrospective analysis of patients in the Mount Sinai Acute GVHD International Consortium (MAGIC) database who underwent HCT between 2014 and 2021 to describe the difference in clinical course of acute GVHD and significance of GVHD biomarkers between Black and White recipients. We used propensity score matching to generate a 1:3 matched cohort of 234 Black patients and 702 White patients with similar baseline characteristics. In the first year after HCT Black patients experienced a higher cumulative incidence of grade III-IV acute GVHD (17% versus 12%, P = 0.050), higher nonrelapse mortality (NRM; 18% versus 12%, P = .009), and lower overall survival that trended toward statistical significance (73% versus 79%, P = .071) compared to White patients. The difference in NRM in the first year was even greater among Black patients who developed GVHD than White patients (24% versus 14%, P = .041). The distribution of low, intermediate, and high MAGIC biomarker scores at the time of treatment was similar across racial groups (P = .847), however, Black patients with high biomarker scores experienced significantly worse NRM than White patients (71% versus 32%, P = .010). Our data indicate that Black patients are at a higher risk of NRM following HCT, primarily from a higher incidence of severe GVHD. Serum biomarkers at treatment initiation can stratify patients for risk of NRM across races, however Black patients with high biomarker scores had a significantly greater NRM risk. These results suggest a need for strategies that mitigate the higher risk for poor GVHD outcomes among Black patients.

Surrogates of Endothelial Injury Predict Survival After Post-transplant Cyclophosphamide.

Post-transplant cyclophosphamide (PT-Cy) is becoming the standard of care for preventing graft-versus-host disease (GVHD) following allogeneic hematopoietic stem cell transplant (alloHCT). Cyclophosphamide is associated with endothelial injury. We hypothesized that the endothelial activation and stress index (EASIX) score, being a marker of endothelial dysfunction, will predict non-relapse mortality (NRM) in alloHCT patients receiving PT-Cy for GVHD prophylaxis. We evaluate the prognostic ability of the hematopoietic cell transplantation-specific comorbidity index (HCT-CI) and EASIX scores, and report other factors influencing survival, in patients with hematologic malignancies undergoing alloHCT and receiving PT-Cy-based GVHD prophylaxis. Adult patients with hematologic malignancies who underwent alloHCT and received PT-Cy for GVHD prophylaxis at the three Mayo Clinic locations were included in this study. We retrospectively reviewed the Mayo Clinic database and the available electronic medical records to determine the patient, disease, and transplant characteristics. An HCT-CI score of ≥3 was considered high. The EASIX score was calculated from labs available between day -28 (of alloHCT) to the day of starting conditioning and analyzed on log2 transformed values. A log2-EASIX score ≥2.32 was considered high. The cumulative incidence of NRM was determined using competing risk analysis, with relapse considered as competing risk. Overall survival (OS) from transplant was determined using Kaplan-Meier and log-rank methods. Cox-proportional hazard method was used to evaluate factors impacting survival. A total of 199 patients were evaluated. Patients with a high log2-EASIX score had a significantly higher cumulative incidence of NRM at 1 year after alloHCT (34.5% versus 12.3%, P = .003). Competing risk analysis showed that a high log2-EASIX score (HR 2.92, 95% CI 1.38 to 6.17, P = .005) and pre-alloHCT hypertension (HR 2.15, 95% CI 1.06 to 4.36, P = .034) were independently predictive of 1 year-NRM. Accordingly, we combined the two factors to develop a composite risk model stratifying patients in low, intermediate, and high-risk groups: 111 (55.8%) patients were considered low-risk, 76 (38.2%) were intermediate and 12 (6%) were high-risk. Compared to patients in the low-risk group, the intermediate (HR 2.38, 95% CI 1.31 to 4.33, P = .005) and high-risk (HR 5.77, 95% CI 2.31 to 14.39, P < .001) groups were associated with a significantly inferior 1-year OS. Multiorgan failure (MOF) was among the common causes of NRM (14/32, 43.8%) particularly among patients with prior pulmonary comorbidities [7 (50%) patients]. Our study shows that EASIX score is predictive of survival after PT-Cy. The novel EASIX-HTN composite risk model may stratify patients prior to transplant. MOF is a common cause of NRM in patients receiving PT-Cy, particularly among patients with pulmonary comorbidities.

Concretizing plan specifications as realizables within the OBO foundry.

BACKGROUND: Within the Open Biological and Biomedical Ontology (OBO) Foundry, many ontologies represent the execution of a plan specification as a process in which a realizable entity that concretizes the plan specification, a "realizable concretization" (RC), is realized. This representation, which we call the "RC-account", provides a straightforward way to relate a plan specification to the entity that bears the realizable concretization and the process that realizes the realizable concretization. However, the adequacy of the RC-account has not been evaluated in the scientific literature. In this manuscript, we provide this evaluation and, thereby, give ontology developers sound reasons to use or not use the RC-account pattern. RESULTS: Analysis of the RC-account reveals that it is not adequate for representing failed plans. If the realizable concretization is flawed in some way, it is unclear what (if any) relation holds between the realizable entity and the plan specification. If the execution (i.e., realization) of the realizable concretization fails to carry out the actions given in the plan specification, it is unclear under the RC-account how to directly relate the failed execution to the entity carrying out the instructions given in the plan specification. These issues are exacerbated in the presence of changing plans. CONCLUSIONS: We propose two solutions for representing failed plans. The first uses the Common Core Ontologies 'prescribed by' relation to connect a plan specification to the entity or process that utilizes the plan specification as a guide. The second, more complex, solution incorporates the process of creating a plan (in the sense of an intention to execute a plan specification) into the representation of executing plan specifications. We hypothesize that the first solution (i.e., use of 'prescribed by') is adequate for most situations. However, more research is needed to test this hypothesis as well as explore the other solutions presented in this manuscript.

Community-engaged artificial intelligence research: A scoping review.

The degree to which artificial intelligence healthcare research is informed by data and stakeholders from community settings has not been previously described. As communities are the principal location of healthcare delivery, engaging them could represent an important opportunity to improve scientific quality. This scoping review systematically maps what is known and unknown about community-engaged artificial intelligence research and identifies opportunities to optimize the generalizability of these applications through involvement of community stakeholders and data throughout model development, validation, and implementation. Embase, PubMed, and MEDLINE databases were searched for articles describing artificial intelligence or machine learning healthcare applications with community involvement in model development, validation, or implementation. Model architecture and performance, the nature of community engagement, and barriers or facilitators to community engagement were reported according to PRISMA extension for Scoping Reviews guidelines. Of approximately 10,880 articles describing artificial intelligence healthcare applications, 21 (0.2%) described community involvement. All articles derived data from community settings, most commonly by leveraging existing datasets and sources that included community subjects, and often bolstered by internet-based data acquisition and subject recruitment. Only one article described inclusion of community stakeholders in designing an application-a natural language processing model that detected cases of likely child abuse with 90% accuracy using harmonized electronic health record notes from both hospital and community practice settings. The primary barrier to including community-derived data was small sample sizes, which may have affected 11 of the 21 studies (53%), introducing substantial risk for overfitting that threatens generalizability. Community engagement in artificial intelligence healthcare application development, validation, or implementation is rare. As healthcare delivery occurs primarily in community settings, investigators should consider engaging community stakeholders in user-centered design, usability, and clinical implementation studies to optimize generalizability.

Novel MAGIC composite scores using both clinical symptoms and biomarkers best predict treatment outcomes of acute GVHD.

ABSTRACT: Acute graft-versus-host disease (GVHD) grading systems that use only clinical symptoms at treatment initiation such as the Minnesota risk identify standard and high-risk categories but lack a low-risk category suitable to minimize immunosuppressive strategies. We developed a new grading system that includes a low-risk stratum based on clinical symptoms alone and determined whether the incorporation of biomarkers would improve the model's prognostic accuracy. We randomly divided 1863 patients in the Mount Sinai Acute GVHD International Consortium (MAGIC) who were treated for GVHD into training and validation cohorts. Patients in the training cohort were divided into 14 groups based on similarity of clinical symptoms and similar nonrelapse mortality (NRM); we used a classification and regression tree (CART) algorithm to create three Manhattan risk groups that produced a significantly higher area under the receiver operating characteristic curve (AUC) for 6-month NRM than the Minnesota risk classification (0.69 vs 0.64, P = .009) in the validation cohort. We integrated serum GVHD biomarker scores with Manhattan risk using patients with available serum samples and again used a CART algorithm to establish 3 MAGIC composite scores that significantly improved prediction of NRM compared to Manhattan risk (AUC, 0.76 vs 0.70, P = .010). Each increase in MAGIC composite score also corresponded to a significant decrease in day 28 treatment response (80% vs 63% vs 30%, P < .001). We conclude that the MAGIC composite score more accurately predicts response to therapy and long-term outcomes than systems based on clinical symptoms alone and may help guide clinical decisions and trial design.

Association of Sociodemographic Factors With Overtriage, Undertriage, and Value of Care After Major Surgery.

Objective: To determine whether certain patients are vulnerable to errant triage decisions immediately after major surgery and whether there are unique sociodemographic phenotypes within overtriaged and undertriaged cohorts. Background: In a fair system, overtriage of low-acuity patients to intensive care units (ICUs) and undertriage of high-acuity patients to general wards would affect all sociodemographic subgroups equally. Methods: This multicenter, longitudinal cohort study of hospital admissions immediately after major surgery compared hospital mortality and value of care (risk-adjusted mortality/total costs) across 4 cohorts: overtriage (N = 660), risk-matched overtriage controls admitted to general wards (N = 3077), undertriage (N = 2335), and risk-matched undertriage controls admitted to ICUs (N = 4774). K-means clustering identified sociodemographic phenotypes within overtriage and undertriage cohorts. Results: Compared with controls, overtriaged admissions had a predominance of male patients (56.2% vs 43.1%, P < 0.001) and commercial insurance (6.4% vs 2.5%, P < 0.001); undertriaged admissions had a predominance of Black patients (28.4% vs 24.4%, P < 0.001) and greater socioeconomic deprivation. Overtriage was associated with increased total direct costs [$16.2K ($11.4K-$23.5K) vs $14.1K ($9.1K-$20.7K), P < 0.001] and low value of care; undertriage was associated with increased hospital mortality (1.5% vs 0.7%, P = 0.002) and hospice care (2.2% vs 0.6%, P < 0.001) and low value of care. Unique sociodemographic phenotypes within both overtriage and undertriage cohorts had similar outcomes and value of care, suggesting that triage decisions, rather than patient characteristics, drive outcomes and value of care. Conclusions: Postoperative triage decisions should ensure equality across sociodemographic groups by anchoring triage decisions to objective patient acuity assessments, circumventing cognitive shortcuts and mitigating bias.

FLT3 inhibitors potentially improve response rates in acute myeloid leukemia harboring t(6;9)(DEK::NUP214): The Mayo Clinic experience.

Not available.

Towards Machine-FAIR: Representing software and datasets to facilitate reuse and scientific discovery by machines.

OBJECTIVE: To use software, datasets, and data formats in the domain of Infectious Disease Epidemiology as a test collection to evaluate a novel M1 use case, which we introduce in this paper. M1 is a machine that upon receipt of a new digital object of research exhaustively finds all valid compositions of it with existing objects. METHOD: We implemented a data-format-matching-only M1 using exhaustive search, which we refer to as M1DFM. We then ran M1DFM on the test collection and used error analysis to identify needed semantic constraints. RESULTS: Precision of M1DFM search was 61.7%. Error analysis identified needed semantic constraints and needed changes in handling of data services. Most semantic constraints were simple, but one data format was sufficiently complex to be practically impossible to represent semantic constraints over, from which we conclude limitatively that software developers will have to meet the machines halfway by engineering software whose inputs are sufficiently simple that their semantic constraints can be represented, akin to the simple APIs of services. We summarize these insights as M1-FAIR guiding principles for composability and suggest a roadmap for progressively capable devices in the service of reuse and accelerated scientific discovery. CONCLUSION: Algorithmic search of digital repositories for valid workflow compositions has potential to accelerate scientific discovery but requires a scalable solution to the problem of knowledge acquisition about semantic constraints on software inputs. Additionally, practical limitations on the logical complexity of semantic constraints must be respected, which has implications for the design of software.

The MAGIC algorithm probability predicts treatment response and long-term outcomes to second-line therapy for acute GVHD.

ABSTRACT: The significance of biomarkers in second-line treatment for acute graft-versus-host disease (GVHD) has not been well characterized. We analyzed clinical data and serum samples at the initiation of second-line systemic treatment of acute GVHD from 167 patients from 17 centers of the Mount Sinai Acute GVHD International Consortium (MAGIC) between 2016 and 2021. Sixty-two patients received ruxolitinib-based therapy, whereas 102 received other systemic agents. In agreement with prospective trials, ruxolitinib resulted in a higher day 28 (D28) overall response Frate than nonruxolitinib therapies (55% vs 31%, P = .003) and patients who received ruxolitinib had significantly lower nonrelapse mortality (NRM) than those who received nonruxolitinib therapies (point estimates at 2-year: 35% vs 61%, P = .002). Biomarker analyses demonstrated that the benefit from ruxolitinib was observed only in patients with low MAGIC algorithm probabilities (MAPs) at the start of second-line treatment. Among patients with a low MAP, those who received ruxolitinib experienced significantly lower NRM than those who received nonruxolitinib therapies (point estimates at 2-year: 12% vs 41%, P = .016). However, patients with high MAP experienced high NRM regardless of treatment with ruxolitinib or nonruxolitinib therapies (point estimates at 2-year: 67% vs 80%, P = .65). A landmark analysis demonstrated that the relationship between the D28 response and NRM largely depends on the MAP level at the initiation of second-line therapy. In conclusion, MAP measured at second-line systemic treatment for acute GVHD predicts treatment response and NRM. The outcomes of patients with high MAP are poor regardless of treatment choice, and ruxolitinib appears to primarily benefit patients with low MAP.

Development and validation of a race-agnostic computable phenotype for kidney health in adult hospitalized patients.

Standard race adjustments for estimating glomerular filtration rate (GFR) and reference creatinine can yield a lower acute kidney injury (AKI) and chronic kidney disease (CKD) prevalence among African American patients than non-race adjusted estimates. We developed two race-agnostic computable phenotypes that assess kidney health among 139,152 subjects admitted to the University of Florida Health between 1/2012-8/2019 by removing the race modifier from the estimated GFR and estimated creatinine formula used by the race-adjusted algorithm (race-agnostic algorithm 1) and by utilizing 2021 CKD-EPI refit without race formula (race-agnostic algorithm 2) for calculations of the estimated GFR and estimated creatinine. We compared results using these algorithms to the race-adjusted algorithm in African American patients. Using clinical adjudication, we validated race-agnostic computable phenotypes developed for preadmission CKD and AKI presence on 300 cases. Race adjustment reclassified 2,113 (8%) to no CKD and 7,901 (29%) to a less severe CKD stage compared to race-agnostic algorithm 1 and reclassified 1,208 (5%) to no CKD and 4,606 (18%) to a less severe CKD stage compared to race-agnostic algorithm 2. Of 12,451 AKI encounters based on race-agnostic algorithm 1, race adjustment reclassified 591 to No AKI and 305 to a less severe AKI stage. Of 12,251 AKI encounters based on race-agnostic algorithm 2, race adjustment reclassified 382 to No AKI and 196 (1.6%) to a less severe AKI stage. The phenotyping algorithm based on refit without race formula performed well in identifying patients with CKD and AKI with a sensitivity of 100% (95% confidence interval [CI] 97%-100%) and 99% (95% CI 97%-100%) and a specificity of 88% (95% CI 82%-93%) and 98% (95% CI 93%-100%), respectively. Race-agnostic algorithms identified substantial proportions of additional patients with CKD and AKI compared to race-adjusted algorithm in African American patients. The phenotyping algorithm is promising in identifying patients with kidney disease and improving clinical decision-making.

Identifying social determinants of health from clinical narratives: A study of performance, documentation ratio, and potential bias.

OBJECTIVE: To develop a natural language processing (NLP) package to extract social determinants of health (SDoH) from clinical narratives, examine the bias among race and gender groups, test the generalizability of extracting SDoH for different disease groups, and examine population-level extraction ratio. METHODS: We developed SDoH corpora using clinical notes identified at the University of Florida (UF) Health. We systematically compared 7 transformer-based large language models (LLMs) and developed an open-source package - SODA (i.e., SOcial DeterminAnts) to facilitate SDoH extraction from clinical narratives. We examined the performance and potential bias of SODA for different race and gender groups, tested the generalizability of SODA using two disease domains including cancer and opioid use, and explored strategies for improvement. We applied SODA to extract 19 categories of SDoH from the breast (n = 7,971), lung (n = 11,804), and colorectal cancer (n = 6,240) cohorts to assess patient-level extraction ratio and examine the differences among race and gender groups. RESULTS: We developed an SDoH corpus using 629 clinical notes of cancer patients with annotations of 13,193 SDoH concepts/attributes from 19 categories of SDoH, and another cross-disease validation corpus using 200 notes from opioid use patients with 4,342 SDoH concepts/attributes. We compared 7 transformer models and the GatorTron model achieved the best mean average strict/lenient F1 scores of 0.9122 and 0.9367 for SDoH concept extraction and 0.9584 and 0.9593 for linking attributes to SDoH concepts. There is a small performance gap (∼4%) between Males and Females, but a large performance gap (>16 %) among race groups. The performance dropped when we applied the cancer SDoH model to the opioid cohort; fine-tuning using a smaller opioid SDoH corpus improved the performance. The extraction ratio varied in the three cancer cohorts, in which 10 SDoH could be extracted from over 70 % of cancer patients, but 9 SDoH could be extracted from less than 70 % of cancer patients. Individuals from the White and Black groups have a higher extraction ratio than other minority race groups. CONCLUSIONS: Our SODA package achieved good performance in extracting 19 categories of SDoH from clinical narratives. The SODA package with pre-trained transformer models is available at https://github.com/uf-hobi-informatics-lab/SODA_Docker.

Amphiregulin, ST2, and REG3α biomarker risk algorithms as predictors of nonrelapse mortality in patients with acute GVHD.

ABSTRACT: Graft-versus-host disease (GVHD) is a major cause of nonrelapse mortality (NRM) after allogeneic hematopoietic cell transplantation. Algorithms containing either the gastrointestinal (GI) GVHD biomarker amphiregulin (AREG) or a combination of 2 GI GVHD biomarkers (suppressor of tumorigenicity-2 [ST2] + regenerating family member 3 alpha [REG3α]) when measured at GVHD diagnosis are validated predictors of NRM risk but have never been assessed in the same patients using identical statistical methods. We measured the serum concentrations of ST2, REG3α, and AREG by enzyme-linked immunosorbent assay at the time of GVHD diagnosis in 715 patients divided by the date of transplantation into training (2004-2015) and validation (2015-2017) cohorts. The training cohort (n = 341) was used to develop algorithms for predicting the probability of 12-month NRM that contained all possible combinations of 1 to 3 biomarkers and a threshold corresponding to the concordance probability was used to stratify patients for the risk of NRM. Algorithms were compared with each other based on several metrics, including the area under the receiver operating characteristics curve, proportion of patients correctly classified, sensitivity, and specificity using only the validation cohort (n = 374). All algorithms were strong discriminators of 12-month NRM, whether or not patients were systemically treated (n = 321). An algorithm containing only ST2 + REG3α had the highest area under the receiver operating characteristics curve (0.757), correctly classified the most patients (75%), and more accurately risk-stratified those who developed Minnesota standard-risk GVHD and for patients who received posttransplant cyclophosphamide-based prophylaxis. An algorithm containing only AREG more accurately risk-stratified patients with Minnesota high-risk GVHD. Combining ST2, REG3α, and AREG into a single algorithm did not improve performance.

Classifying early infant feeding status from clinical notes using natural language processing and machine learning.

The objective of this study is to develop and evaluate natural language processing (NLP) and machine learning models to predict infant feeding status from clinical notes in the Epic electronic health records system. The primary outcome was the classification of infant feeding status from clinical notes using Medical Subject Headings (MeSH) terms. Annotation of notes was completed using TeamTat to uniquely classify clinical notes according to infant feeding status. We trained 6 machine learning models to classify infant feeding status: logistic regression, random forest, XGBoost gradient descent, k-nearest neighbors, and support-vector classifier. Model comparison was evaluated based on overall accuracy, precision, recall, and F1 score. Our modeling corpus included an even number of clinical notes that was a balanced sample across each class. We manually reviewed 999 notes that represented 746 mother-infant dyads with a mean gestational age of 38.9 weeks and a mean maternal age of 26.6 years. The most frequent feeding status classification present for this study was exclusive breastfeeding [n = 183 (18.3%)], followed by exclusive formula bottle feeding [n = 146 (14.6%)], and exclusive feeding of expressed mother's milk [n = 102 (10.2%)], with mixed feeding being the least frequent [n = 23 (2.3%)]. Our final analysis evaluated the classification of clinical notes as breast, formula/bottle, and missing. The machine learning models were trained on these three classes after performing balancing and down sampling. The XGBoost model outperformed all others by achieving an accuracy of 90.1%, a macro-averaged precision of 90.3%, a macro-averaged recall of 90.1%, and a macro-averaged F1 score of 90.1%. Our results demonstrate that natural language processing can be applied to clinical notes stored in the electronic health records to classify infant feeding status. Early identification of breastfeeding status using NLP on unstructured electronic health records data can be used to inform precision public health interventions focused on improving lactation support for postpartum patients.