Mapping of transdiagnostic neuropsychiatric phenotypes across patients in two general hospitals

Abstract

Background –

Multidimensional transdiagnostic phenotyping systems are increasingly important to neuropsychiatric phenotyping, particularly in translational research settings. The relationship the NIMH Research Domain Criteria (RDoC) multidimensional approach to psychopathology and nonpsychiatric diagnoses has not been studied at scale but is relevant to those caring for neuropsychiatric illness in medical and surgical settings.

Methods –

We applied the CQH Dimensional Phenotyper natural language processing tool to estimate NIMH RDoC domain associated symptoms of individuals admitted to nonpsychiatric wards at each of two large academic general hospitals over an eight-year period. We compared patterns in individual domain symptom burden, as well as a new pooled unidimensional measure, by primary medical and surgical diagnosis.

Results –

Analysis included 227,243 patients from hospital 1 of whom 68,793 (30.3%) had a prior psychiatric history and 220,213 patients from hospital 2 of whom 50,818 (23.1%) had a prior psychiatric history. The distribution of RDoC symptom burdens over primary diagnosis was similar across hospital sites and differed significantly across primary medical or surgical diagnosis. The effect of primary medical or surgical diagnosis was larger than that of prior psychiatric history on RDoC symptom burden.

Conclusion –

RDoC based neuropsychiatric symptom burden estimated from general hospital patients’ clinical documentation is more strongly associated with the primary hospital medical or surgical diagnosis than it is with the presence of a previous psychiatric history. The bidirectional role of psychiatric and somatic illness warrants further study through the lens of transdiagnostic phenotyping.

Introduction

Recent initiatives in psychiatry emphasize the utility of characterizing psychopathology in terms of dimensional features of symptomatology, rather than through dichotomous diagnostic categories.^1–4 Dimensional approaches offer the possibility of studying differences within and across existing categorical diagnoses which might in turn track modifiable biology, treatment response, or long-term outcome.^5–8

Whereas the consult psychiatry literature includes extensive work on co-occurring categorical medical and psychiatric diagnosis^9–16 less is known about the distribution of dimensional psychopathology across medical and surgical presentations. The existing literature with traditional diagnostic categories suggests that co-morbid psychiatric disorders are common among those admitted to the general hospital with medical or surgical complaints.^17–22 This comorbidity is not just common but also consequential, as co-morbid psychopathology portends a more complicated hospital course when present.^23–25

The National Institute of Mental Health’s (NIMH) Research Domain Criteria (RDoC) is a dimensional approach to neuropsychiatric illness which aims to study brain diseases based on six dimensions each capturing a spectrum of symptoms and can be rooted in brain circuits and biology.^1,2,26 The original RDoC domains were negative valence systems, positive valence systems, cognitive systems, systems for social processes, and arousal/regulatory systems, with a sixth sensorimotor domain newly added.^27,28 While batteries for assessment of individual domains have been developed, typically comprising some combination of self-report measures and neuropsychological tests, they are not feasible for large-scale administration to hospitalized patients, particularly if the intention is to capture multiple domains. Additional these newly developed batteries cannot be linked to large existing data sets. Recognizing these twin limitations we have previously shown that symptom burdens associated with original five NIMH RDoC domain can be estimated from electronic health records (EHRs) using natural language processing (NLP) as implemented in the CQH Dimensional Phenotyper.^6,29,30 In addition to validation against clinical assessment, prior work has shown that these transdiagnostic phenotypes have genetic correlates, associate with death by suicide, risk stratify for new onset dementia, and predict healthcare utilization in children and adults.^6,31–35

In this study, we applied the CQH Dimensional Phenotyper NLP methodology to estimate symptom burden by RDoC domain among individuals hospitalized on medical and surgical units at two academic general hospitals. This approach allows large-scale assessment of dimensional neuropsychopathology among non-psychiatric populations, extending prior work that focused on individual psychiatric diagnosis in particular medical or surgical populations.³⁵

Materials and methods

Cohort and Sample

This study used cohorts drawn from two separate Massachusetts academic medical centers. Both hospital cohort included all adults admitted to non-psychiatric services between January 1, 2005 and December 31, 2013 – a period after universal use of electronic documentation but before wide dissemination of the NIMH RDoC framework. For patients who had more than one admission during the observation period, a single encounter was selected at random. For all individuals included in the study coded claims data, demographics, and discharge summary documentation were extracted from the hospitals’ EHRs and used to generate an i2b2 datamart.³⁶ Extracted data included sex, race, age on admission, classification of primary insurance as public or private, primary discharge diagnosis, presences or absence of a psychiatric diagnostic code prior to admission, and hospital discharge summary note. In order to produce systematic grouping of admissions by the primary discharge diagnostic codes were collapsed to the top level of the Healthcare Utilization Project (HCUP) Clinical Classification Software (CCS) hierarchy.³⁷ The chapter categories included obstetrical and perinatal, congenital, dermatologic, endocrine, injury, musculoskeletal, urological, neoplastic, hematologic, gastrointestinal, pulmonary, circulatory, infectious, neurological, psychiatric, and other as categories of primary hospital diagnosis.

Natural Language Processing RDoC Estimation

Hospital discharge summaries were used to derive estimated RDoC (eRDoC) symptom scores using a previously published open source NLP tool.²⁹ In brief, this RDoC domain estimation software uses a set of terms relevant to each of the five domain. These domain specific term lists were developed through an iterative collaboration between expert clinicians and unsupervised machine learning. This iterative process merged knowledge of clinical documentation and the original intent of the RDoC framework authors as captured in the NIMH published definitions of the domains and in collaboration with the NIMH RDoC working group.^6,29 Although the NIMH RDoC source materials are freely available, to briefly recapitulate, within the RDoC framework the negative valiance domain aims to captures experiences like fear, anxiety, and loss which are aversive;³⁸ the positive valiance domain aims to captures reward seeking, consummatory behavior, and reward or habit learning which are positive motivational situations;³⁸ the cognitive domain aims to captures cognitive processes;^34,38 the social process domain aims to captures perception and interpretation of others’ actions and responses in interpersonal settings;³⁸ and finally, the arousal and regulatory domain aims to capture systems of homeostatic regulation for systems like energy balance and sleep and neural system activation matched to various contexts.³⁸

In the scoring software the domain-specific term lists are used to derive symptom burden estimates for each of the five RDoC domain. The score is calculated as the percent of domain-specific terms appearing in any given note. For example, if 50 of 100 possible terms appear in a discharge summary, that discharge summary scores 50/100 or 0.5 for that domain. This process is repeated for each of the five separate domains such that a given clinical document is converted into five numerical estimates of RDoC symptom burden. The software used for this scoring is freely available as open source for use and inspection.¹ Unlike many forms of clinical NLP this, application of this trained model does not require an intricate preprocessing step for lexical normalization and tokenization; instead, the final trained model includes specific words and phrases for direct use. This approach facilitates reuse, replication, and deployment by shifting the complexity into engineering the domain features and away from the point of application to any given chart. The eRDoC symptom burden scores derived by the tool have been linked to utilization, death by suicide, symptom change over hospitalization, and genetics. As these domains are independent, we extended prior work reporting on domains individually by also examining the sum of all five domains, to produce a single unifactorial symptom burden score which ranges from 0–500% instead of 0–100%, in an approach motivated by prior research suggesting a unidimensional structure to psychiatric disorders.³

Design and Analysis

The included patient cohorts were characterized using summary statistics as appropriate to the variable. Unifactorial eRDoC scores were plotted across hospital sites by HCUP CCS diagnostic chapter to provide a readily interpretable summary of trends over primary hospital diagnostic groups. Differences in unifactorial eRDoC score across diagnostic groups were tested for using analysis of variance and, where significant, characterized using Tukey’s post hoc test. Pearson product-moment correlation of mean unifactorial eRDoC by diagnostic group was used to establish similarity of diagnostic group trends across the two hospital sites. To characterize the relative distribution of symptom burden within each RDoC domain across CCS diagnostic groups while controlling for demographic variables, we fit a linear regression per RDoC domain predicted by patient sex, race, use of public insurance, z scored age, note length, and hospital site, then calculated the mean residual for each domain by primary diagnosis chapter.

To characterize the distribution of RDoC symptoms at the individual level we mapped a subsample of individual patients through the five-dimensional estimated eRDoC space. To do so, we generated a balanced subsample of 1,000 patients per HCUP CCS diagnostic chapter then visualized that balanced subsample using t-Distributed Stochastic Neighbor Embedding (t-SNE), an approach to dimensionality reduction and visualization developed for machine learning.³⁹ To quantify the degree of medical diagnostic segregation achieved using the five RDoC dimensions we calculated five-dimensional Euclidean distance among all patients in the balanced mapping subsample and calculated the portion of cases in which the nearest neighbor (the individual with the shortest Euclidean distance from each patient) had the same primary diagnostic category. All analysis utilized R version 4.

The study protocol was approved by the Partners HealthCare Human Research Committee. No participant contact was required in this study, which relied on secondary use of data produced by routine clinical care, allowing waiver of requirement for informed consent.

Results

The cohort included 227,243 patients from hospital 1 of whom 105,193 (46.3%) were male and 181,147 (79.7%) were white and 220,213 patients from hospital 2 of whom 78,273 (35.5) were male and 158958 (72.2%) were white (Table 1). As admissions to psychiatry were excluded and diagnostic classification was based on primary diagnosis, psychiatric diagnosis was rare, accounting for only 1.4% (3,135) of admissions to hospital 1 and 0.6% (1,423) of admissions to hospital 2; however, psychiatric history among this cohort of general hospital patients was common with 30.3% (68,793) of patients having a prior psychiatric diagnosis at hospital 1 and 23.1% (50,818) having a prior psychiatric history at hospital 2. At hospital 1, cardiovascular admissions were most common and accounted for 21.9% (49,874) of all admissions, whereas at hospital 2, obstetrical and prenatal admissions were most common accounting for 23.7% (52,215) of all admissions.

Table 1.

Summary characteristics of the two hospital cohorts

	Hospital 1	Hospital 2
n	227243	220213
Sex = M (%)	105193 (46.3)	78273 (35.5)
Race = white (%)	181147 (79.7)	158958 (72.2)
Insurance = public (%)	94102 (41.4)	77929 (35.4)
Age (mean (SD))	55.66 (19.12)	52.19 (18.96)
Any Psychiatric History = True (%)	68793 (30.3)	50818 (23.1)
Primary Diagnosis Group (%)
Circulatory	49874 (21.9)	38736 (17.6)
Congenital	2307 (1.0)	1243 (0.6)
Dermatologic	3806 (1.7)	1955 (0.9)
Endocrine	8420 (3.7)	6943 (3.2)
Gastrointestinal	18241 (8.0)	13672 (6.2)
Hematologic	1223 (0.5)	1447 (0.7)
Infectious	646 (0.3)	432 (0.2)
Injury	30878 (13.6)	22563 (10.2)
Musculoskeletal	18395 (8.1)	14081 (6.4)
Neoplastic	34914 (15.4)	40008 (18.2)
Neurological	7586 (3.3)	5786 (2.6)
OB/Perinatal	23127 (10.2)	52215 (23.7)
Other	4736 (2.1)	2992 (1.4)
Psychiatric	3135 (1.4)	1423 (0.6)
Pulmonary	9575 (4.2)	7319 (3.3)
Urological	10380 (4.6)	9398 (4.3)
eRDoC Scores
Unifactorial (mean (SD))	30.65 (17.53)	28.93 (21.63)
Arousal   Regulatory (mean (SD))	7.77 (4.53)	7.84 (6.03)
Cognitive (mean (SD))	4.67 (3.75)	4.24 (3.87)
Negative (mean (SD))	6.06 (4.01)	5.09 (4.21)
Positive (mean (SD))	8.67 (5.34)	8.40 (6.80)
Social process (mean (SD))	3.49 (3.14)	3.36 (3.15)

The distribution of unifactorial eRDoC scores by hospital site and primary diagnostic group are shown in Figure 1. The median unifactorial eRDoC score within each diagnostic group was highly correlated across the two hospital sites (r(14) = .90, p < .001; Figure 1) suggesting stability of effect across hospitals. In combined analysis of the two hospital cohorts, unifactorial eRDoC scores differed significantly across primary diagnostic group (F(15, 447440) = 4775.27, p < .001, η²_p = .14). In post hoc testing of differences among diagnostic 113 of 120 pairs differed significantly (Supplemental Table 1). Without regard to primary CCS diagnostic group, patients with a preadmission psychiatric history had an average unifactorial eRDoC score of 32.6 whereas those without a prior psychiatric history had a significantly lower (t(203298.22) = −56.00, p < .001, d = −0.19) average unifactorial eRDoC score of 28.8. When considered simultaneously, the effect of prior psychiatric history remained significant (F(1, 447439) = 3801.05, p < .001, η²_p < .01) but small compared to that of primary diagnostic group (F(15, 447439) = 4640.70, p < .001, η²_p =.13). That is, the estimated symptom burden quantified as summed eRDoC domain scores is associated with both the primary medical or surgical diagnosis (η²_p =.13) and also, to a lesser extent, the presence or absence of formal psychiatric history (η²_p <.01). Median unifactorial eRDoC score per CCS diagnostic group was highly correlated across patients with and without psychiatric history (r(14) = .99, p < .001) consistent with an equivalent disruption of cumulative symptom burden across primary medical or surgical diagnosis regardless of prior psychiatric diagnosis.

Figure 1.

Distribution of unifactorial eRDoC scores by primary diagnosis HCUP chapter across the two hospital cohorts.

The average residual eRDoC for each of the five domains after controlling for demographic factors is shown by primary diagnostic group in Figure 2. Figure 2 facilitates comparisons within symptom domain across primary diagnostic group such that the distribution of a given symptom burden over primary diagnosis can be characterized. Among the many possible comparisons which could be made the pattern of more arousal symptom burden among endocrine and neurological patients, higher cognitive and social symptom burden among neurological and psychiatric patients, lower negative symptom burden among obstetrical and musculoskeletal patients, and higher positive symptom burden among psychiatric, but not neurologic, patients stand out as larger magnitudes.

Figure 2.

Average per domain residual eRDoC after controlling for age, sex, race, public insurance, note length, and admission hospital by primary diagnosis HCUP chapter.

Figure 3 illustrates the distribution in five dimensional RDoC space of a random subset of the full cohort stratified by primary diagnosis represented in two dimensions using tSNE. Figure 3A includes the full subset and shows significant overlap in RDoC space of the primary diagnostic categories; however, small regions of diagnostic concentration standout as plot areas that are monochromatic. Figure 3B isolates the two diagnostic groups (psychiatric and obstetrics) which stood out in Figure 3A and highlights a pattern of both large areas of overlap and small higher density areas of relative case segregation. Figure 4 quantifies proximity of patients in each diagnostic category by reporting the percent of closest, by Euclidean distance, neighbors in five dimensional RDoC space who share the same primary admission diagnosis. This approach asks the question - how often does the patient with the most similar five eRDoC phenotypes also have the same primary medical or surgical diagnostic chapter as by CCS captured? Obstetrical patients are most segregated in eRDoC space with 46% of patients having a fellow obstetrical patient as a nearest neighbor vs. the 6.2% expected by chance alone; when viewed in the context of Figure 1 that commonality is likely the absence of symptom burden. Obstetrical patients were followed by psychiatric (22.8%), endocrine (15.6%), neurological (14.8%) and musculoskeletal (13%) as the five diagnostic groups most likely to have a nearest neighbor in eRDoC space in the same primary diagnostic group. Pulmonary (6.2%), injury (6.8%), urological (7.1%), cancer (7.2%), and cardiovascular (7.8%) patients were least likely to have a nearest eRDoC neighbor in the same diagnostic group.

Figure 3.

t-SNE derived two-dimensional representation of five dimensional eRDoC space color coded by primary diagnostic group showing all included groups in A and then limiting to only two of the diagnostic groups (obstetrics and psychiatry) in B.

Figure 4.

Percent of patients in each diagnostic group whose nearest neighbor in five dimensional eRDoC was in the same diagnostic group.

Discussion

Transdiagnostic multidimensional approaches are an increasingly important approach to psychiatric nosology.⁴⁰ In this study of admission to two large general hospitals spanning 447,456 individuals, we found that transdiagnostic neuropsychiatric symptom domains distinguish primary hospital diagnostic categories reliably but not precisely and, in turn, that symptom burden is unevenly distributed over primary nonpsychiatric diagnostic categories. Unifactorial symptom burden was lowest in obstetrical patients and highest in psychiatric patients and the pattern of distributions between these diagnostic groups was common across the two hospitals. This result provides a comparable quantification of cumulative neuropsychiatric symptom burden across medical and surgical inpatients presentations and captures a stably replicated of that interrelationship between behavioral symptoms and somatic illness at two medical centers.

The five eRDoC symptom domains showed variable patterns across primary diagnostic categories and the primary medical diagnosis explained a larger share of variability in RDoC symptom burden than presence of formal psychiatric history. This differential symptom distribution and increased importance of underlying primary illness points toward opportunities to deliver targeted interventions and improve on existing proactive psychiatric interventions which rely heavily on psychiatric history.^41–44 At the individual patient level, obstetrical patients were most likely to be adjacent to another obstetrical patient, whereas pulmonary patients were least likely to be situated adjacent to a fellow pulmonary patient in five-dimensional symptom space. Given the pooled distribution of symptoms observed across those groups it seems likely that obstetrical admissions are characterized by their shared paucity of documented symptoms. Further work is warranted to characterize both the extent to which dimensional neuropsychiatric phenotyping distinguishes among diagnostic groups and, perhaps more importantly, identifies clinical significant subgroups within primary diagnosis.³⁵

Although the present study extends the consult psychiatry literature on dimensional phenotyping, the result must be interpreted in light of limitations inherent in the approach. The RDoC based neuropsychiatric symptom burden estimation tool used in this work predates the recent release of the RDoC motor symptom domain. Although this lack of the sixth domain is unfortunate the lack of the additive sixth domain does not undermine the estimation of the original five domains and it allows the use of a previously validated tool with multiple prior publications. In order to avoid influence from provider knowledge of RDoC itself the sample is limited to individuals hospitalized before 2013. This approach also avoids the complexity arising from the transition to ICD 10 and DSM 5.⁴⁵ Finally, both hospital sites included are academic medical centers and thus, although symptom burdens are stable over primary diagnosis across the two included tertiary care centers, the extent to which relative these patterns would generalize to other settings is unknown.

Drawing on a large EHR sample generated at two general hospitals, the present result enables comprehensive comparison of psychiatric symptom burden across primary diagnostic categories and comparison of symptom composition within primary presenting diagnosis. This extends the extensive prior work within psychosomatic medicine on medical-psychiatric comorbidity based on categorical diagnosis of each and points to many future directions for subsequent research at the intersection of somatic illness and psychiatric symptom burden. Although the NIMH RDoC framework is a notable approach to dimensional phenotyping it not the only system for dimensional phenotyping.^4,46 Future work aimed at characterizing the distribution of psychopathology over patients in the general hospital could extend this RDoC based work to personality focused systems,^47,48 systems which focuses on minimally differentiated affective content,^49,50 data driven high dimensional approaches,⁵¹ or recapitulate prior efforts to engineer specific algorithms to estimate prespecified phenotypes with an eye toward the unidimensional p factor.^3,6,29,52 In addition to exploring alternative transdiagnostic phenotyping systems future work is needed to characterize the role of comorbid categorical diagnosis – that is the role of medical illness present during an admission other than the primary diagnosis for the admission. Regardless of the particular direction in dimensional phenotyping taken and independent of the hope that dimensional approaches will lead to deeper biological insights and threptic breakthroughs the practical efficiencies of dimensional approaches are sufficient to warrant greater attention in future psychosomatic research: the number of categorical psychiatric and categorical somatic diagnostic entities is large and the unique combinations among these multiplicatively larger so as to be impractical for individual study as distinct clinical entities. Those combinations are, however, the purview of the consulting psychiatrist aiming to bridge physical and mental health. Dimensional neuropsychiatric phenotypes offer a tractable approach to a comprehensive clinical evidence base which is nevertheless tailored to primary medical or surgical illness.

Conclusion

We find consistent patterns of NIMH RDoC domains by medical or surgical diagnosis between two academic medical centers. Most notably for consulting psychiatrists working in general hospital settings, the primary medical or surgical hospital diagnosis explained more of the variation in RDoC symptom burdens than prior psychiatric diagnosis did. This result suggests, through the lens of a contemporary multidimensional phenotyping system, that the impact of comorbid psychiatric and medical illness may be greater than the sum of the parts.