Phenotyping Down syndrome: discovery and predictive modelling with electronic medical records

Abstract

Background: Individuals with Down syndrome (DS) have a heightened risk for various co-occurring health conditions, including congenital heart disease (CHD). In this two-part study, electronic medical records (EMRs) were leveraged to examine co-occurring health conditions among individuals with DS (Study 1) and to investigate health conditions linked to surgical intervention among DS cases with CHD (Study 2).

Methods: De-identified EMRs were acquired from Vanderbilt University Medical Center and facilitated creating a cohort of N = 2282 DS cases (55% females), along with comparison groups for each study. In Study 1, DS cases were one-by-two sex and age matched with samples of case-controls and of individuals with other intellectual and developmental difficulties (IDDs). The phenome-disease association study (PheDAS) strategy was employed to reveal co-occurring health conditions in DS versus comparison groups, which were then ranked for how often they are discussed in relation to DS using the PubMed database and Novelty Finding Index. In Study 2, a subset of DS individuals with CHD [N = 1098 (48%)] were identified to create longitudinal data for N = 204 cases with surgical intervention (19%) versus 204 case-controls. Data were included in predictive models and assessed which model-based health conditions, when more prevalent, would increase the likelihood of surgical intervention.

Results: In Study 1, relative to case-controls and those with other IDDs, co-occurring health conditions among individuals with DS were confirmed to include heart failure, pulmonary heart disease, atrioventricular block, heart transplant/surgery and primary pulmonary hypertension (circulatory); hypothyroidism (endocrine/metabolic); and speech and language disorder and Alzheimer's disease (neurological/mental). Findings also revealed more versus less prevalent co-occurring health conditions in individuals with DS when comparing with those with other IDDs. Findings with high Novelty Finding Index were abnormal electrocardiogram, non-rheumatic aortic valve disorders and heart failure (circulatory); acid-base balance disorder (endocrine/metabolism); and abnormal blood chemistry (symptoms). In Study 2, the predictive models revealed that among individuals with DS and CHD, presence of health conditions such as congestive heart failure (circulatory), valvular heart disease and cardiac shunt (congenital), and pleural effusion and pulmonary collapse (respiratory) were associated with increased likelihood of surgical intervention.

Conclusions: Research efforts using EMRs and rigorous statistical methods could shed light on the complexity in health profile among individuals with DS and other IDDs and motivate precision-care development.

Keywords: Down syndrome; behavioural phenotypes; intellectual difference; methodology in research.

Figure 1.

Study 1 examined novel conditions co-occurring with Down syndrome (DS). (a) The phenome-disease association study (PheDAS) analysis identified electronic medical record (EMR) phecodes that are significantly associated with DS in our cohort. (b) The clinical novelty of each identified phecode was assessed by calculating its Novelty Finding Index (NFI), a relative novelty measure that compares the reliability of the association to how ‘well studied’ the phecode is on PubMed.

Figure 2.

Study 2 focused on subjects with both Down syndrome (DS) and congenital heart disease (CHD); predictive modelling was used to identify conditions prevalent prior to heart surgery in this population. (a) Surgery (case) subjects were matched to non-surgery subjects according to age and sex; electronic medical records (EMRs) after the case subject’s surgery age were removed. (b) EMRs are mapped to phecodes and aggregated; three different predictive models were trained to classify subjects as surgery or non-surgery based on these aggregated phecodes. (c) An explainable modelling method [local interpretable model-agnostic explanations (LIME)] is used to estimate the contribution of each phecode to an individual’s surgery prediction; these are averaged across subjects to estimate the generalised contribution of each phecode to a heart surgery outcome. SVM, support vector machine; RF, random forest; MLP, multi-layer perceptron.

Figure 3.

Health conditions related to the likelihood of surgical intervention among Down syndrome cases with congenital heart disease based on model-based explanatory predictors from the best-performing multi-layer perceptron classifier.