Search

Objective Attention networks learn an intelligent weighted averaging mechanism over a series of entities, providing increases to both performance and interpretability. In this article, we propose a novel time-aware transformer-based network and compare it to another leading model with similar characteristics. We also decompose model performance along several critical axes and examine which features contribute most to our model’s performance. Materials and methods Using data sets representing patient records obtained between 2017 and 2019 by the Kaiser Permanente Mid-Atlantic States medical system, we construct four attentional models with varying levels of complexity on two targets (patient mortality and hospitalization). We examine how incorporating transfer learning and demographic features contribute to model success. We also test the performance of a model proposed in recent medical modeling literature. We compare these models with out-of-sample data using the area under the receiver-operator characteristic (AUROC) curve and average precision as measures of performance. We also analyze the attentional weights assigned by these models to patient diagnoses. Results We found that our model significantly outperformed the alternative on a mortality prediction task (91.96% AUROC against 73.82% AUROC). Our model also outperformed on the hospitalization task, although the models were significantly more competitive in that space (82.41% AUROC against 80.33% AUROC). Furthermore, we found that demographic features and transfer learning features which are frequently omitted from new models proposed in the EMR modeling space contributed significantly to the success of our model. Discussion We proposed an original construction of deep learning electronic medical record models which achieved very strong performance. We found that our unique model construction outperformed on several tasks in comparison to a leading literature alternative, even when input data was held constant between them. We obtained further improvements by incorporating several methods that are frequently overlooked in new model proposals, suggesting that it will be useful to explore these options further in the future.

Objective To construct and publicly release a set of medical concept embeddings for codes following the ICD-10 coding standard which explicitly incorporate hierarchical information from medical codes into the embedding formulation. Materials and Methods We trained concept embeddings using several new extensions to the Word2Vec algorithm using a dataset of approximately 600,000 patients from a major integrated healthcare organization in the Mid-Atlantic US. Our concept embeddings included additional entities to account for the medical categories assigned to codes by the Clinical Classification Software Revised (CCSR) dataset. We compare these results to sets of publicly released pretrained embeddings and alternative training methodologies. Results We found that Word2Vec models which included hierarchical data outperformed ordinary Word2Vec alternatives on tasks which compared naïve clusters to canonical ones provided by CCSR. Our Skip-Gram model with both codes and categories achieved 61.4% normalized mutual information with canonical labels in comparison to 57.5% with traditional Skip-Gram. In models operating on two different outcomes, we found that including hierarchical embedding data improved classification performance 96.2% of the time. When controlling for all other variables, we found that co-training embeddings improved classification performance 66.7% of the time. We found that all models outperformed our competitive benchmarks. Discussion We found significant evidence that our proposed algorithms can express the hierarchical structure of medical codes more fully than ordinary Word2Vec models, and that this improvement carries forward into classification tasks. As part of this publication, we have released several sets of pretrained medical concept embeddings using the ICD-10 standard which significantly outperform other well-known pretrained vectors on our tested outcomes.

Background People living with human immunodeficiency virus (HIV; PLWH) have a markedly elevated anal cancer risk, largely due to loss of immunoregulatory control of oncogenic human papillomavirus infection. To better understand anal cancer development and prevention, we determined whether recent, past, cumulative, or nadir/peak CD4+ T-cell count (CD4) and/or HIV-1 RNA level (HIV RNA) best predict anal cancer risk. Methods We studied 102 777 PLWH during 1996–2014 from 21 cohorts participating in the North American AIDS Cohort Collaboration on Research and Design. Using demographics-adjusted, cohort-stratified Cox models, we assessed associations between anal cancer risk and various time-updated CD4 and HIV RNA measures, including cumulative and nadir/peak measures during prespecified moving time windows. We compared models using the Akaike information criterion. Results Cumulative and nadir/peak CD4 or HIV RNA measures from approximately 8.5 to 4.5 years in the past were generally better predictors for anal cancer risk than their corresponding more recent measures. However, the best model included CD4 nadir (ie, the lowest CD4) from approximately 8.5 years to 6 months in the past (hazard ratio [HR] for Conclusions Our results are consistent with anal cancer promotion by severe, prolonged HIV-induced immunosuppression. Nadir and cumulative CD4 may represent useful markers for identifying PLWH at higher anal cancer risk.

Background Agents from the integrase inhibitor (INSTI) therapeutic class only are recommended as initial therapy for most patients with HIV. Clinicians now face a decision when treating ART-experienced patients on non-INSTI regimens: continue current therapy or switch to INSTI. Multiple factors may be considered in this decision: clinician/patient preference, comorbidities, tolerability, and resistance history. The objective of this analysis was to examine patient factors associated with currently taking an INSTI-based regimen. Methods We used data from the DC Cohort, a longitudinal observational cohort of patients receiving HIV care at 14 clinics between 2011–2018. Participants in the sample had ≥ 1 encounter between 4/1/17 and 3/1/18, were aged ≥ 18 years and were ART experienced. Participants were classified as currently, previously, or never on an INSTI. Independent variables included demographics, clinical characteristics, alcohol/tobacco use, HBV/HCV status and HIV-related variables (recent CD4 and HIV RNA, presence of resistance mutations). Multivariable multinomial logistic regression was used to identify factors associated with INSTI use status. Results Among 4584 participants (58.2% aged 50+ years; 69.4% male; 2.5% transgender; 80.3% Black; 36% MSM), most (65.0%) were current INSTI users; however, a sizeable proportion (28.3%) were never users and 6.7% were former users. Current and previous INSTI users were more likely to have a major NRTI, NNRTI or PI mutation compared with never users (see Table 1). Transgender participants (compared with males), were less likely to be current (vs. never) users (adjusted odds ratio (aOR) 0.48, 95% CI 0.32, 0.72). Younger participants (18–24 vs 50+ years) were more likely current users (aOR 1.90, 95% CI 1.18, 3.06), as were Hispanic participants (aOR 1.39, 95% CI 1.05, 1.84). Conclusion The majority of active DC Cohort participants were using INSTI-based therapy. Transgender and older individuals were less likely to be on INSTIs, indicating that they are more likely to be on PI-based or NNRTI-based therapy or not on therapy. Further research should explore whether this is detrimental for long-term HIV outcomes in these patient groups. Additionally, these results suggest resistance history as an important driver of INSTI prescription. Disclosures All authors: No reported disclosures.

Background Among persons living with HIV (PLWH) with type 2 diabetes mellitus (DM) there is limited research on the effect of DM control on CD4 count. Current guidelines recommend that PLWH with DM maintain a hemoglobin A1c (HbA1c) Methods We used data from the DC Cohort, a longitudinal observational cohort of patients receiving HIV care at 14 clinics between 2011–2018. Participants with DM on an ongoing antiretroviral regimen with ≥1 year of follow-up, ≥2 HbA1c results, and ≥2 CD4 count results were included. Participants were compared based on the most recent HbA1c result categorized into one of three control levels control: strict, HbA1c < 7.5%; moderate, HbA1c between 7.5–9.0%; and uncontrolled, HbA1c >9.0%. All statistical tests were performed within the framework of the linear mixed-effects (LME) model. The rates of increase in CD4 count by DM control were compared using an LME model with random slopes and random intercepts, adjusted for sex, BMI, nadir CD4, a history of AIDS, or cancer diagnosis. Results Among 554 participants (median age 53.5; 70.8% male; 82.7% Black), there were 5,138 total CD4 count measurements. In unadjusted analysis, participants with moderate or uncontrolled HbA1c had higher mean CD4 counts over the follow-up period than those with strict HbA1c control (strict: 690 cells/μL, moderate: 712 cells/μL uncontrolled: 711 cells/μL; P = 0.0156 strict vs. moderate, 0.049 strict vs. uncontrolled). All DM control groups had a similar temporal increase over time in CD4 count (P = 0.46). In multivariate analysis, only moderate vs. strict control showed a significant difference in CD4 count (mean difference=18.1; P = 0.02). Results showed CD4 count change was not affected by the duration of HIV diagnosis or diabetes diagnosis. See Table 1 for additional results. Conclusion PLWH and DM with moderate HbA1c control had higher CD4 counts than those with strict HbA1c control and similar CD4 counts compared with those with uncontrolled HbA1c levels, while the rate of increase in CD4 count was similar in the three groups. These results show that moderate DM control may benefit CD4 count, which should be considered when revising DM control guidelines for PLWH. Disclosures All authors: No reported disclosures.

The risk of developing tuberculosis among persons infected with human immunodeficiency virus (HIV) has been reported to be extremely high in the first 3 months after initiation of highly active antiretroviral therapy (HAART), with rates ranging from 1300 to 1700 per 100000 person-years in developed countries [1, 2] and from 10700 to 23000 per 100000 person-years in developing countries [2, 3]. Tuberculosis diagnosed after HAART initiation may represent previously undiagnosed prevalent disease, newly acquired tuberculosis (ie, acquired since HAART initiation), or progression of subclinical tuberculosis that was not recognized prior to initiation of antiretroviral therapy (so-called tuberculosis unmasking) [4]. A subset of persons in the latter group may have immune reconstitution inflammatory syndrome, with a severe and/or pronounced clinical presentation [5–7]. With continued use of HAART, tuberculosis risk decreases over time, although the risk remains higher than in HIV-seronegative persons [1–3]. Tuberculosis rates shortly after HAART initiation could potentially be decreased if screening for active tuberculosis prior to HAART initiation were optimized. In the United States, it is recommended that persons infected with HIV undergo screening for latent Mycobacterium tuberculosis infection at the time of HIV diagnosis, and annually thereafter if they are at high risk for exposure to tuberculosis [8]. However, the guidelines do not address screening prior to HAART initiation—for either latent M. tuberculosis infection or active disease. Screening for tuberculosis prior to HAART initiation is not routinely performed in the United States and Canada [9, 10]. Risk factors for developing tuberculosis shortly after HAART initiation are incompletely understood. Predisposing factors in developed countries include low CD4+ lymphocyte count prior to HAART initiation and HIV transmission risk factors of heterosexual sex and injection drug use [1, 2]. In developing countries, risk factors include low CD4+ lymphocyte count, younger age, and male sex [2]. However, in a study from South Africa, only low CD4+ lymphocyte count in the 4 months prior to tuberculosis diagnosis was significantly associated with increased tuberculosis risk during antiretroviral therapy [3]. Low CD4+ lymphocyte count prior to antiretroviral therapy initiation also appears to be associated with “tuberculosis unmasking” [5, 7, 11]. Unmasked tuberculosis can become acid-fast bacillus smear-positive and therefore perhaps more readily transmitted to others than subclinical disease [7]. In one study, tuberculosis diagnosed in AIDS patients during the first 3 months of antiretroviral therapy was also associated with a 3-fold higher mortality rate than when diagnosed before or more than 3 months after antiretroviral therapy initiation [12]. The authors suggested that patients with tuberculosis unmasking have a higher mortality rate due to delays in tuberculosis diagnosis and initiation of antituberculosis therapy. A better understanding of the risk factors for developing tuberculosis in the first few months after antiretroviral therapy initiation—the objective of the present analysis—can help inform screening strategies to target patients at highest risk, both before and shortly after antiretroviral therapy initiation.