Introduction

Advanced heart failure (ADHF) occurs in almost 5% of all hospitalized patients and represents a clinical intersection of therapeutic alternatives that can be utilized depending on the patient’s treatment goals and preexisting comorbidities1. Durable mechanical circulatory support devices such as left ventricular assist devices (LVADs) have emerged as a vital therapeutic option for patients with ADHF2, offering significant improvements in quality of life (QoL)2, acting as a bridge to heart transplantation (BTT) for eligible candidates and adequate circumstances3, or as a destination therapy (DT). For these reasons, LVADs have become increasingly prevalent4. For instance, in 2021, a total of 2464 LVADs were implanted in the United States, and hospitalizations for an implant of these devices are among the most expensive inpatient stays covered by public insurance coverage (Medicare)5, ranging from 120 000 to 270 000 USD in 2014 and 250 000 and 280 000 USD in 20195,6. Although health expenditure for newer devices is lower6, longer inpatient stays during index admissions and postoperative recurrent hospitalizations represent an important economic burden. While LVADs have demonstrated remarkable success in prolonging survival, it is essential to recognize that these devices are not without severe complications, such as neurological dysfunction (ND), that could lead to even higher health expenditure, morbidity, and mortality. The Interagency Registry for Mechanically Assisted Circulatory Support (INTERMACS) defines ND as any abnormality related to the central or peripheral nervous system that is attributable to the use of mechanical circulatory support (MCS)7. This includes a broad spectrum of neurological events, ranging from temporary neurological impairment (delirium, seizures, or transient ischemic attacks [TIAs]) without structural damage evidence (type 3) to subclinical structural evidence of injury (type 2) to overt ischemic or hemorrhagic strokes (type 1). Prior investigations have indicated that patients undergoing LVAD implantation often present with a prior history of stroke, small vessel cerebral disease (SVCD), manifested as white matter lesions, cerebral microbleeds or brain atrophy, TIAs, and brain tumors8,9,10. The interplay between pre-existing neurological conditions or injuries and the subsequent development of ND in LVAD therapy remains an area of active research. Retrospective registry studies and clinical trials have been valuable sources of information to gauge the incidence, risk factors, and therapeutic alternatives of these complications11,12,13,14,15,16,17. These studies have offered valuable insights, particularly for patients developing strokes or TIAs as a complication of LVAD therapy, even during the third-generation continuous-flow (CF) LVAD era12,14. Nevertheless, there is scant evidence regarding the incidence and impact of other types of ND, perioperative identification, and postoperative resource allocation (frequency of self-care training sessions, inpatient rehabilitation disposition, etc.)18,19,20,21, partly due to a lack of structural and functional quantitative neuroimaging evidence of brain damage before LVAD implants due to the use of preimplant cardiac devices (intra-aortic balloon pump, catheter-based cardiac assist devices), which are MRI incompatible. Nonetheless, evidence suggests that prior neurological damage has an important role in delirium development in postoperative courses22,23,24, and MRI-based studies suggest an inconclusive association between the incidence of postoperative delirium in non-cardiac surgery and pre-surgical brain quantitative markers25,26,27,28,29. Hence, due to the ongoing use of LVADs as a treatment option for ADHF and the impact of ND, particularly delirium, on inpatient mortality and morbidity during LVAD index admissions, there is a need for reliable biomarkers that can inform clinicians about management strategies for non-vascular ND during heart transplantation and LVAD workup. Leveraging contemporary neuroimaging methods and preliminary data from existing studies, this research aimed to investigate the volume of lateral ventricles as a potential marker of brain parenchymal loss to identify patients at higher risk of experiencing postoperative delirium (type-3 ND) during durable LVAD insertion index admissions.

Methods

Design, patients, and clinical settings

We designed a prospective pilot study to evaluate global and regional brain parenchymal loss performance to discriminate patients developing postoperative delirium (type-3 ND) during the first 30 days after CF-LVAD implantation. We aimed to study brain volumetric markers previously associated with postoperative and ICU-related delirium22,23,24,25,26,27,28,29,30 and subcortical structures considered to be vulnerable to global and regional atrophy due to shared and common modifiable and non-modifiable risk factors in LVAD patients, such as advanced age31,32,33, obesity34,35, cardiovascular risk31,36, prior vascular pathology (small vessel cerebral disease, large vessel occlusion, lacunar strokes, etc.)36. In addition, we contemplated the reliability of targeted anatomical structures, considering that they were analyzed using clinical CT scan images. Therefore, we selected the expansion of lateral ventricles (LV) as the primary marker and the decrease of cortical gray matter (cGM), cerebral white matter (WM), thalamus (TL), and basal ganglia (BG) volumes as secondary markers. We sought to study the association of lateral ventricle fraction (LVF) with postoperative delirium incidence and its performance as a biomarker.

A flow diagram depicting the enrollment process is reported in Fig. 1. The study was conducted on 166 ADHF patients with LVAD support at our institution from March 2020 to July 2023. The protocol and consent forms were approved by the Institutional Review Board (H-45532 and H-51123). Patients were screened from the Heart Transplant Program using the following inclusion criteria: (1) Age > 18 years; (2) Advanced heart failure stage NYHA class IV, stage D; (3) Left ventricular ejection fraction (LVEF) ≤ 25%; (4) Inotropic dependency; (5) cardiac index (CI) < 2.2 L/min/m2 while on inotropes; (6) Medical Review Board and insurance approval for durable LVAD insertion. Patients diagnosed with any form of dementia (Alzheimer’s disease, vascular dementia, dementia with Lewy bodies, Parkinson’s disease-related dementia, frontotemporal dementia, and mixed dementia) and those not agreeing to participate in the study were excluded from the study. Subjects who underwent an explantation of the device during the study period were eliminated from the research. The primary outcome was defined using the five key features of the American Psychiatric Association’s Diagnostic and Statistical Manual, fifth edition (DSM-5)37. The DSM-5 criteria for delirium include disturbance in attention and awareness, change in cognition, acute onset and fluctuating course, and evidence of an underlying cause.

Fig. 1
figure 1

Flowchart of the study enrollment. Abbreviations: OHT: orthopic heart transplantation; CT: computed tomography.

Full size image

Ethics approval

All procedures conducted in studies involving human participants adhered to the ethical standards set forth by the institutional review board at Baylor College of Medicine, with approval granted under protocol numbers H-45532 and H-51123. This research complied with the principles outlined in the 1964 Helsinki Declaration and its subsequent amendments.

Consent

In strict adherence to the ethical guidelines outlined in the Federal Policy for the Protection of Human Subjects, also known as ‘The Common Rule’ (45 CFR 46), and the provisions of the Health Insurance Portability and Accountability Act (HIPAA), informed consent was diligently obtained from each participant enrolled in the study, ensuring compliance with regulatory frameworks governing human subjects’ protection and privacy regulations.

Surgical procedure, clinical evaluation, and management

All patients underwent a comprehensive clinical evaluation for heart transplant approval before LVAD insertion. This evaluation included a transthoracic Doppler ultrasound of the heart, a bilateral Doppler ultrasound of the leg and carotid arteries, a pulmonary function test, and a standardized psychosocial assessment conducted by a licensed practitioner. The assessment included the Stanford Integrated Psychosocial Assessment for Transplant (SIPAT)38, which rates patients from 1 to 120 on readiness and illness management, social support, psychological stability, psychopathology, lifestyle, and the impact of substance abuse. Additionally, the Alcohol Use Disorders Identification Test and CAGE questionnaires were administered if there was a history of substance abuse. A specialized nurse practitioner provided LVAD self-care training for patients. All patients discontinued antiplatelet agents 5 days before the surgery, and warfarin was started on day 5 postoperatively to achieve an International Normalized Ratio (INR) of 2–3. To operationalize the components of DSM-5 into daily evaluations of patients, we utilized the Confusion Assessment Method in the ICU (CAM-ICU)39 by a trained critical care physician, following institutional protocol. The CAM-ICU is a validated tool specifically designed to detect delirium in ICU patients and aligns well with the DSM-5 criteria since it includes the following components: acute onset or fluctuating course, inattention, altered level of consciousness, and disorganized thinking. Delirium was confirmed through chart review based on dual confirmation by two different assessors using a validated method, achieving an overall agreement of 80% and a kappa of 0.3040. Sedation and anesthesia for the surgery were administered using a standardized combination of propofol, fentanyl, dexmedetomidine, isoflurane, and sevoflurane, according to the center’s cardiovascular surgery anesthesia management protocol. Patients underwent median sternotomy and mechanical circulatory support (MCS) with cardiopulmonary bypass and anterograde del Nido cardioplegia. Postoperatively, all patients were transferred to the cardiovascular ICU recovery area, where non-pharmacological and pharmacological management of postoperative delirium was initiated. A detailed protocol for managing delirium in the ICU is provided in the online Supplementary Information.

CT scan acquisition and processing

Clinical non-contrast CT scans of the brain were acquired 1 to 5 days before the LVAD implant procedure, using a Siemens® SOMATOM Definition Edge CT scanner. Coronal images were obtained with a slice thickness of 3 mm, and an acquisition matrix of 512 × 594 pixels was used. The reconstructed images had a diameter of 188 mm and pixel spacing of approximately 0.316 mm in both directions. The scan employed a tube voltage (KVP) of 120 kV and an exposure time of 1000 ms. Radiation exposure parameters indicated a CTDIvol of 43.41 mGy and an estimated dose saving of 22%. The Hounsfield unit (HU) scale was used to rescale the images, and the images were reconstructed with the convolutional kernel Hf38s to allow better gray-white matter differentiation. Coronal images were visually inspected to evaluate artifacts and the overall quality of the images. After being evaluated and selected, de-identified DICOM files were converted to NIfTI files with MRIcron (v1.0.2). The scans were clipped to 0–80 HU; then, we performed symmetric diffeomorphic registration to a stroke-population-specific CT template41 (Supplementary Information). Brain segmentation was done with a neural network (SynthSeg v2.0) capable of segmenting brain scans of different contrasts and resolutions by using synthetic images generated on the fly and exposing the neural network to random parameters of the generative model, such as resolution, orientation, artifacts, or contrasts42 (Supplementary Information). Segmentations were visually inspected and manually corrected when appropriate with ITK-SNAP (v4.0.1). Left and right segmentations and volumes obtained from corrected segmentations were fused into one single mask and calculation. Merged segmentations were modeled using spherical harmonization (SPHARM)43 functions and represented as point distribution models obtaining a parametrically spherical mesh and a surface mesh, and then these models underwent rigid alignment using user-defined reference points to improve the correspondence among SPHARM representations, as implemented in SlicerSALT (v4.0.1). SPHARM average models were displayed in ParaView (v5.13) (Supplementary Information).

Statistical analysis

Descriptive data were analyzed using the Mann-Whitney U or t-test test for continuous variables and chi-squared tests for categorical variables. Data were reported in frequencies and percentages (%) for qualitative data and median and interquartile ranges (IQR) or mean and standard deviation (SD) for quantitative data, as appropriate. Brain metrics not displaying a normal distribution were assessed for different transformations (see Supplementary Figure S1), and the transformation was subsequently used for analyses. We adjusted brain volumes to total intracranial volume (TIV) to account for head size and provide fractional measures of each structure (LVF, cGMF, WMF, TLF, BGF). Then we transformed them into percent to allow better interpretability. Regression models were employed to assess the association of neuroimaging markers with postoperative delirium, using age as a covariate to address confounding. Interaction terms between age and brain volume fractions (LVF, cGMF, WMF, TLF, BGF) were included in subsequent binomial regression models to evaluate the influence of aging on brain atrophy. Due to the lack of prior evidence of postoperative delirium incidence in LVAD patients, we performed simulation-based post hoc sample size and power estimations based on different hypothetical conditions including the simulation of similar conditions proposed in our study (ρ1 = 0.45, ρ2 = 0.55, a = 0.05, event ratio 1:3.8 [prevalence 33%]), and assuming that a difference of 0.08 between an area under the curve (AUC) of 0.74 and 0.82 to be clinically relevant in this context (see Supplementary Figure S2). To assess the statistical significance of morphological differences potentially attributed to age, TIV, and comorbidities (including prior stroke history) while considering spatial smoothness, spatial correlation, and low-dimensional representation of functional shape responses, we employed a multivariate functional shaped data analysis (MFSDA) based on a local Wald-type test statistic. The interactions between age and comorbidities were examined in patients with and without delirium, and the results are presented on an average mesh with false discovery rate P-values (P-FDR). Diagnostic performance metrics and Wald test chi-squared (χ2) were calculated with corresponding 95% confidence intervals (CI). Additionally, F1 scores were calculated to evaluate precision and recall on imbalanced data considering the age groups, and precision-recall curves were plotted. Receiving operating curves (ROC) and the area under the curve (AUC) were calculated and plotted to visualize the performance of LVF in predicting postoperative delirium. We replicated the analyses in a subgroup of patients ≥ 60 years old to assess the association and performance of LVF in identifying postoperative delirium. Additionally, we calculated E-values to provide estimates of the minimum strength of association that an unmeasured confounder would need to have with both the exposure and the outcome to fully explain away the observed association. This approach provides quantitative estimations to contextualize the potential impact of unmeasured clinical factors or perioperative variables on the development of postoperative delirium. The statistical analyses were performed on StataMP (v18.0), R (4.3.2), and SlicerSALT (v4.0.1).

Results

Main features of the sample population

The demographic and clinical characteristics of the study population are displayed in Table 1. The study population consisted of 166 individuals, divided into those who experienced delirium (n = 56) and those who did not (n = 110). The average age was higher in the delirium group compared to the non-delirium group. The prevalence of type-2 diabetes, and chronic kidney disease (CKD) stage III or higher and white matter hypodensities in NCCT was higher in in the delirium group. Medical history of cortical stroke in the posterior cerebral artery was significantly more common in patients experiencing delirium. Chronic liver disease was less frequently present in the delirium group. Other characteristics showed no significant differences between the groups (Table 1).

Table 1 Clinical and demographic characteristics of the study population.
Full size table

Surgical procedure, delirium, and postoperative course

All patients underwent surgery through a median sternotomy and cardiopulmonary bypass (CPB). There were no statistically significant differences in intraoperative variables; however, postoperative outcomes regarding hospital admission length and self-care LVAD training were longer in patients who experienced postoperative delirium (Table 2).

Table 2 Perioperative characteristics and operative and admission outcomes.
Full size table

Regional atrophy in patients with and without postoperative delirium

Patients who developed postoperative delirium demonstrated a pronounced expansion of their lateral ventricles, reflecting a significant increase in brain atrophy. This condition was illustrated by an elevated mean lateral ventricular fraction (LVF) (log) (Fig. 2a). In addition to this ventricular enlargement, these patients also exhibited notable reductions in the volume of various types of parenchymal tissue. Specifically, there was a marked decrease in the volume of cortical gray matter (Fig. 2b), white matter (Fig. 2c), basal ganglia (Fig. 2d), and the thalamus (Fig. 2e). All these changes serve as indicators of increased brain atrophy. The differences in brain tissue volumes were statistically significant and ranged from moderate to large, with Cohen’s d estimates between 0.5 and 0.9. Among these, the most substantial changes were observed in the LVF. Interestingly, when focusing on patients aged over 60, only the LVF (log) (Fig. 2f) and white matter (Fig. 2g) remained significantly different from those without delirium, while no such differences were evident for gray matter (Fig. 2h), basal ganglia (Fig. 2i), and the thalamus (Fig. 2j). Even in these older patients, the difference in LVF was particularly pronounced, maintaining its categorization as a large effect.

Fig. 2
figure 2

The plots illustrate the comparisons of five brain volume metrics: LVF, cGMF, WMF, BGF, and TLF in all patients of the cohort (a through e) and in patients > 60 years old (f through j) with and without postoperative delirium. Each plot highlights the differences in mean and median values between the two groups, showing higher ventricular expansion in all patients and those > 60 years old. Abbreviations: LVF, lateral ventricular fraction; cGMF, cortical gray matter fraction; WMF, white matter fraction; BGF, basal ganglia fraction; TLF, thalamus fraction.

Full size image

Association between postoperative delirium with age and regional brain atrophy

To understand and estimate the influence of age on subcortical atrophy and the occurrence of postoperative delirium, we fitted different linear models using age as a predictor of lateral ventricles expansion. We noted that patients with and without postoperative delirium displayed a significant association of age with LVF (log) (Fig. 3a) and LVF (%) (Fig. 3b). However, no linear associations were found in patients with or without postoperative delirium between age and LVF (Fig. 3c and d).

Fig. 3
figure 3

Scatter fit plots demonstrate the relationship between age and log-transformed LVF (a) and LVF percentage (b) in the entire patient cohort. Additionally, they show the relationship between age and log-transformed LVF (c) and LVF percentage (d) in patients over 60 years old. Significant associations are observed between age and LVF in the overall cohort; however, no significant associations are found in patients over the age of 60. Abbreviations: LVF, lateral ventricles fraction. Estimates and parameter descriptions.

Full size image

Ventricular shape morphology in patients experiencing postoperative delirium

The average mesh representation of the fused lateral ventricles, reconstructed using Spherical Harmonic Analysis for the Representation of Morphology (SPHARM) presented in Fig. 4. The statistical shape model of fused lateral ventricles met the spherical assumption and revealed significant insights into the relationship between ventricular morphology and age in patients experiencing postoperative delirium. The multivariate functional shape data analysis (MFSDA) modeling demonstrated that there was a significant relationship between years of age and ventricular expansion only in patients who experienced postoperative delirium. Older patients exhibited greater ventricular enlargement and shape deformation than younger patients. Vertex variations in the shape model highlighted specific regions of the ventricles, predominantly at the body (central region), where these deformations were most pronounced, with FDR P-values (< 0.05) displayed in red on an average mesh, providing a detailed map of age-related morphological changes (Fig. 4). Patients without postoperative delirium did not show any significant association with vertex variations.

Fig. 4
figure 4

Average mesh representations of the fused left ventricle (LV) are displayed in six standard anatomical views: superior (a), anterior (b), left lateral (c), inferior (d), posterior (e), and right lateral (f). Red clusters in the figure indicate significant surface expansion in patients with postoperative delirium compared to those without, following a Bonferroni correction for multiple comparisons. These clusters highlight the association of expansion with age, suggesting that brain atrophy is more pronounced in older patients with postoperative delirium. Open-source software SlicerSALT (https://salt.slicer.org/) and Paraview (https://www.paraview.org/) were used to generate Fig. 4.

Full size image

Association between brain volumetrics, clinical factors, and postoperative delirium

Table 3 demonstrates a significant correlation between various brain volumetric markers and the occurrence of postoperative delirium. Notably, LVF demonstrated a strong association with postoperative delirium in the overall patient cohort and patients over 60 years old. Similarly, cGMF, WMF, and TLF were associated with the incidence of postoperative delirium. However, the just LVF (log) remained significantly associated in patients > 60 years old. While BGF and TLF showed significant associations in unadjusted models, these associations lost significance after adjustment (Table 3). The models aimed to evaluate the influence of age on LVF and postoperative delirium indicated similar results (Table S1). The interaction between age and LVF did not yield significant results in the overall patient cohort (OR 1.30; 95% CI 0.16–4.34; P = 0.099). Similar findings were observed in patients over 60 years old (OR 0.98; 95% CI 0.86–1.50; P = 0.741) (Table S1).

Further analysis to identify the association between modifiable and non-modifiable risk factors and postoperative delirium is presented in the online supplemental information (see Supplementary Table S2 and Table S3). We noticed that a history of type-2 diabetes was significantly associated with postoperative delirium in the unadjusted and adjusted model in all patients (Table S2), and this association remained significant in the unadjusted and adjusted model of patients > 60 years old (see Supplementary Table S3). Similarly, the unadjusted model for CKD stage ≥ III history showed a significant association with postoperative delirium. This association remained significant in the adjusted model (Table S2). In patients > 60 years old, history of CKD stage ≥ III was not significantly associated in unadjusted and adjusted models (see Supplementary Table S3).

The evaluation of unmeasured confounders using E values revealed that an unmeasured variable would need to be associated with the outcome (postoperative delirium) with an OR between 4.82 and 5.46 to invalidate the association between LVF (log) and postoperative delirium. Furthermore, an OR between 5.64 and 5.68 would be necessary to completely explain away the association of LVF (log) with postoperative delirium in patients over 60 years of age (Table S4).

Table 3 Association between brain volumetrics and postoperative delirium.
Full size table

Baseline LVF in identifying patients at risk for postoperative delirium

We used age-adjusted logistic models to assess the effectiveness of baseline LVF in identifying patients at risk for postoperative delirium. Our results showed that LVF (log) performed well in all cohorts (Fig. 5a) and in patients over 60 years old (Fig. 5c), with minimal differences in AUC. Similarly, LVF (%) adjusted for age effectively distinguished patients with postoperative delirium in the overall group (Fig. 5b) and those over 60 (Fig. 5d), also showing minimal AUC differences. Additional metrics, including sensitivity and specificity, are available in the online supplemental information (see Supplementary Table S3).

Fig. 5
figure 5

ROC curves showing the classification, precision, and recall performance of age-adjusted LVF as a predictor for postoperative delirium in all the cohorts employing LVF (log) (a) and LVF (%) (b), and LVF (log) (c) and LVF (%) (d) in patients age > 60 years old.

Full size image

Discussion

Our study quantitatively analyzed clinical non-contrast CT scans of the brain in ADHF patients undergoing LVAD insertion to determine the association between preexisting brain atrophy and postoperative delirium, and its adequacy in identifying at-risk patients before surgery. Our findings suggest that pronounced brain parenchymal loss, indicated by a higher LVF at the time of LVAD implantation, is associated with an increased incidence of postoperative delirium, regardless of age. Age is one of the most significant factors driving delirium after both cardiac44,45,46 and non-cardiac surgery47,48. Additionally, we observed that the age-adjusted LVF reasonably predicts the incidence of postoperative delirium. WM, cortical GM, and subcortical GM atrophy were significantly lower in patients experiencing postoperative delirium, consistent with previous reports with similar findings22,23,26,27. However, these differences were slightly smaller compared to the changes observed with the LV. We propose that the expansion of the LV observed in patients with delirium is substantial and strongly associated with the condition. Although this expansion cannot be precisely attributed to either gray or white matter, its significance is evident. Our study did not reveal statistically significant associations between parenchymal volumetric estimations (cGMF, WMF, BGF, TLF) and postoperative delirium in patients over 60. However, there is an indication of an atrophying effect, which may be partially mitigated by employing synthetic images derived from low-resolution CT scans to quantify volumes (segmentation of CSF tends to be more precise in comparison with gray and white matter structures). Notably, in the entire cohort, cGMF, WMF, and TLF were associated with postoperative delirium. This suggests that cortical and subcortical atrophy, consistent with the expansion of the lateral ventricles, is a significant factor in the development of delirium.

Furthermore, the role of age on LV expansion, as shown by morphological differences in age-shape associations and the incidence of delirium, appears to be prominent. This indicates that not only is the overall change in cerebral mass tissue relevant, but also the pattern of morphological shifts accompanying this process in the development of postoperative delirium.

These findings resembled the behavior of age over the structural integrity of brain structures previously reported in healthy and diseased individuals31,36, suggesting that the baseline structural integrity of the brain is highly related to the incidence of ND in ADHF patients undergoing LVAD insertion. Even though age was the main driver of brain atrophy and postoperative delirium, we suspect that additional factors such as prior cumulative cerebrovascular injury (sustained brain hypoperfusion, cardiorenal syndrome, long-standing hypertension, etc.) over the CNS and the surgical procedure itself (systemic inflammation, pump use, transient hypo- hypervolemia, and prolonged ventilation) are two additional factors involved in the development of postoperative delirium in these circumstances. This could be partially supported by prior studies using multimodal neuroimaging describing the effect of high cardiovascular risk and pre-established macro and microstructural brain injury30,33,36.

We believe that patients in our study with more pronounced brain atrophy were in the later stages of SVCD49. Preoperative MRI studies in patients undergoing cardiac surgery have shown that larger ventricular volumes, new ischemic lesions, and white matter disease (deep subcortical white matter and periventricular hyperintensities) are linked to postoperative delirium21,22,50. This suggests that in the earlier stages of ischemic cardiomyopathy before LVAD implantation, the brain might have structural abnormalities that increase the risk of postoperative delirium due to CNS injuries, such as blood-brain barrier (BBB) disruption51,52,53. Studies on patients undergoing on-pump cardiac and non-cardiac surgery have shown that postoperative BBB disruption, assessed through CSF enhancement (hyperintense acute reperfusion marker or HARM) in FLAIR sequences or contrast extravasation subtraction, correlates significantly with lesions on diffusion-weighted images54 and cognitive function51. Additionally, BBB disruption, indicated by CSF enhancement, has been associated with advanced age, cerebral microbleeds, and brain atrophy55, which are part of the SVCD spectrum.

Findings from the present study might support the use of clinical non-contrast brain CT scans enhanced by the generation of synthetic images to assess the risk of ND in patients undergoing LVAD implantation. Due to the varying availability and LVAD-compatibility of brain MRI, non-contrast CT scans present a promising alternative for quantitative analyses by (1) providing objective evidence of neurological risk for postoperative management and (2) utilizing an imaging exam commonly performed in ADHF patients to rule out cerebrovascular disease, a “strongly adverse factor” for heart transplant according to Medicare coverage rules56. Notably, a significant proportion of patients in our study had central lines or probes inserted on the right side of the neck before surgery, preventing a complete internal carotid artery (ICA) Doppler ultrasound to identify carotid stenosis, a recognized postoperative delirium risk factor50,57, and a test included in the stroke, transient ischemic attack (TIA), and carotid dissection workup. Similarly, pre-implant neuropsychological evaluations58,59,60,61 that could aid in estimating delirium risk cannot be easily or reliably performed in ventilated, sedated, or extremely fatigued patients, which could occur in up to 30–40% of patients, particularly those with INTERMACS profile I to II62. Identifying at-risk patients for postoperative delirium is crucial for anticipating strategies to manage cognitive demands after LVAD implantation, improving LVAD self-care training proficiency, identifying overlapping ND, shortening length of stay (LOS), and preventing complications. We believe that the quantification of brain volumetrics using synthetic-based imaging pipelines, along with the application of normalized metrics (such as z-scores) derived from multicenter validation studies, holds significant promise for implementing this intervention in patients undergoing LVAD workup.

Although these observations hold significant clinical implications, we must acknowledge the limitations of our study. First, this is a single-center observational study without a control group and with a relatively small sample size, which introduces selection bias, temporality issues, and reverse causality, potentially challenging the establishment of causal relationships, and even though the estimations of unmeasured confounding showed that the findings regarding lateral ventricular expansion would need considerably large effect sizes from unmeasured confounders, such as, acute renal dysfunction, oxygenation or hemodynamic instability to discard the observations, the conclusions should be carefully interpreted. Importantly, slightly different acquisition protocols of brain imaging from different eras could disrupt the localization of anatomical structures and result in large differences in volumetric estimations. Additionally, the sample size for the subanalyses in patients > 60 years old was smaller than the original cohort, therefore the estimations provided by those models are not powered for the study’s main hypothesis. Another issue with imaging acquisition arises from the varying times and days when CT scans were obtained, as lateral ventricle volume changes throughout the day and under different circumstances. Additionally, we did not perform neuropsychological testing to obtain quantitative metrics for cognitive performance before surgery. Using non-contrast CT brain scans with 3-mm resolution is also suboptimal for conventional quantitative neuroimaging analyses due to partial low resolution and problematic gray-white matter differentiation in some individuals, potentially omitting subtle structural details. We employed a 3D deep learning method to segment brain scans; however, 3D segmentation models might be sensitive to the misalignment of medical images and require high computational resources. These limitations could be mitigated using a 2D segmentation method, which decreases image coregistration dependencies and computational load. To address these limitations, we employed a robust deep-learning network to register prior brain lesions in MRI and CT brain scans and generate synthetic images with considerable precision. Furthermore, the enrollment of patients in our study was inclusive enough to provide a representative sample with characteristics partially similar to those in multicenter registries and clinical trials3,4.

Insights from this preliminary work pave the way for further studies on the structural integrity of the CNS in patients with acute decompensated heart failure, both young and old, as well as those with ischemic cardiomyopathy (ICM) and non-ischemic cardiomyopathy (NICM). Initially, robust deep learning algorithms can be employed to analyze neuroimaging data acquired for clinical purposes, such as ruling out cerebrovascular disease, to study ND in a population with diverse risk levels. Additionally, segmentation and generation of synthetic scans from native non-contrast clinical CT scans of the brain could provide valuable insights into the association between CNS structural integrity and postoperative outcomes, offering alternatives to improve resource utilization. Prospective multicentric studies with larger sample sizes from different centers and varied CT acquisition protocols are needed to address the lack of generalizability of these findings, validate the performance of this biomarker, and provide reliable normative metrics in this population. These iterations will help to answer whether the preoperative characterization of brain structural integrity might offer a new understanding of the current ND classification by INTERMACS or help to tailor individual ND risks and management strategies (sedation protocols, training planning, etc.) for postoperative delirium, TIAs, and strokes. Consequently, these observations might provide valuable insights for designing meaningful endpoints in clinical trials for durable LVADs that could partially rely on the degree of CNS structural integrity before intervention, considering highly relevant contextual factors such as age, cardiomyopathy etiology, sex, and INTERMACS profile.

Conclusions

The use of clinical non-contrast CT brain imaging is a promising tool for assessing the risk of postoperative delirium. New-generation processing pipelines can enrich and augment imaging data, offering a feasible alternative when contrast high-resolution CT or conventional MRI is contraindicated. The findings suggest that significant brain parenchymal loss, not attributable to gray or white matter atrophy at the time of LVAD implantation, may drive the incidence of neurological dysfunction in patients with ADHF undergoing LVAD. However, these results should be replicated in a multicenter validation study. These and future insights are considered to be crucial for designing future studies to assess the risk of neurological dysfunction in patients undergoing durable MCS.