Tracing Tomorrow: young people's preferences and values related to use of personal sensing to predict mental health, using a digital game methodology

BACKGROUND

Use of personal sensing to predict mental health risk has sparked interest in adolescent psychiatry, offering a potential tool for targeted early intervention.

OBJECTIVES

We investigated the preferences and values of UK adolescents with regard to use of digital sensing information, including social media and internet searching behaviour. We also investigated the impact of risk information on adolescents' self-understanding.

METHODS

Following a Design Bioethics approach, we created and disseminated a purpose-built digital game (www.tracingtomorrow.org) that immersed the player-character in a fictional scenario in which they received a risk assessment for depression Data were collected through game choices across relevant scenarios, with decision-making supported through clickable information points.

FINDINGS

The game was played by 7337 UK adolescents aged 16-18 years. Most participants were willing to personally communicate mental health risk information to their parents or best friend. The acceptability of school involvement in risk predictions based on digital traces was mixed, due mainly to privacy concerns. Most participants indicated that risk information could negatively impact their academic self-understanding. Participants overwhelmingly preferred individual face-to-face over digital options for support.

CONCLUSIONS

The potential of digital phenotyping in supporting early intervention in mental health can only be fulfilled if data are collected, communicated and actioned in ways that are trustworthy, relevant and acceptable to young people.

CLINICAL IMPLICATIONS

To minimise the risk of ethical harms in real-world applications of preventive psychiatric technologies, it is essential to investigate young people's values and preferences as part of design and implementation processes.

Anemia, hemoglobin concentration and cognitive function in the Longitudinal Ageing Study in India-Harmonized Diagnostic Assessment of Dementia (LASI-DAD) and the Health and Retirement Study

Background

In India, anemia is widely researched in children and women of reproductive age, however, studies in older populations are lacking. Given the adverse effect of anemia on cognitive function and dementia this older population group warrants further study. The Longitudinal Ageing Study in India - Harmonized Diagnostic Assessment of Dementia (LASI-DAD) dataset contains detailed measures to allow a better understanding of anaemia as a potential risk factor for dementia.

Method

2,758 respondents from the LASI-DAD cohort, aged 60 or older, had a complete blood count measured from venous blood as well as cognitive function tests including episodic memory, executive function and verbal fluency. Linear regression was used to test the associations between blood measures (including anemia and hemoglobin concentration (g/dL)) with 11 cognitive domains. All models were adjusted for age and gender with the full model containing adjustments for rural location, years of education, smoking, region, BMI and population weights.Results from LASI-DAD were validated using the USA-based Health and Retirement Study (HRS) cohort (n=5720) to replicate associations between blood cell measures and global cognition.

Results

In LASI-DAD, we showed an association between anemia and poor memory (p=0.0054). We found a positive association between hemoglobin concentration and ten cognitive domains tested (β=0.041-0.071, p<0.05). The strongest association with hemoglobin was identified for memory-based tests (immediate episodic, delayed episodic and broad domain memory, β=0.061-0.071, p<0.005). Positive associations were also shown between the general cognitive score and the other red blood count tests including mean corpuscular hemoglobin concentration (MCHC, β=0.06, p=0.0001) and red cell distribution width (RDW, β =-0.11, p<0.0001). In the HRS cohort, positive associations were replicated between general cognitive score and other blood count tests (Red Blood Cell, MCHC and RDW, p<0.05).

Conclusion

We have established in a large South Asian population that low hemoglobin and anaemia are associated with low cognitive function, therefore indicating that anaemia could be an important modifiable risk factor. We have validated this result in an external cohort demonstrating both the variability of this risk factor cross-nationally and its generalizable association with cognitive outcomes.

Clinical Prompt Learning With Frozen Language Models

When the first transformer-based language models were published in the late 2010s, pretraining with general text and then fine-tuning the model on a task-specific dataset often achieved the state-of-the-art performance. However, more recent work suggests that for some tasks, directly prompting the pretrained model matches or surpasses fine-tuning in performance with few or no model parameter updates required. The use of prompts with language models for natural language processing (NLP) tasks is known as prompt learning. We investigated the viability of prompt learning on clinically meaningful decision tasks and directly compared this with more traditional fine-tuning methods. Results show that prompt learning methods were able to match or surpass the performance of traditional fine-tuning with up to 1000 times fewer trainable parameters, less training time, less training data, and lower computation resource requirements. We argue that these characteristics make prompt learning a very desirable alternative to traditional fine-tuning for clinical tasks, where the computational resources of public health providers are limited, and where data can often not be made available or not be used for fine-tuning due to patient privacy concerns. The complementary code to reproduce the experiments presented in this work can be found at https://github.com/NtaylorOX/Public_Clinical_Prompt.

Integrated methylome and phenome study of the circulating proteome reveals markers pertinent to brain health

Characterising associations between the methylome, proteome and phenome may provide insight into biological pathways governing brain health. Here, we report an integrated DNA methylation and phenotypic study of the circulating proteome in relation to brain health. Methylome-wide association studies of 4058 plasma proteins are performed (N = 774), identifying 2928 CpG-protein associations after adjustment for multiple testing. These are independent of known genetic protein quantitative trait loci (pQTLs) and common lifestyle effects. Phenome-wide association studies of each protein are then performed in relation to 15 neurological traits (N = 1,065), identifying 405 associations between the levels of 191 proteins and cognitive scores, brain imaging measures or APOE e4 status. We uncover 35 previously unreported DNA methylation signatures for 17 protein markers of brain health. The epigenetic and proteomic markers we identify are pertinent to understanding and stratifying brain health.

Personalised treatment for cognitive impairment in dementia: development and validation of an artificial intelligence model

BACKGROUND

Donepezil, galantamine, rivastigmine and memantine are potentially effective interventions for cognitive impairment in dementia, but the use of these drugs has not been personalised to individual patients yet. We examined whether artificial intelligence-based recommendations can identify the best treatment using routinely collected patient-level information.

METHODS

Six thousand eight hundred four patients aged 59-102 years with a diagnosis of dementia from two National Health Service (NHS) Foundation Trusts in the UK were used for model training/internal validation and external validation, respectively. A personalised prescription model based on the Recurrent Neural Network machine learning architecture was developed to predict the Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) scores post-drug initiation. The drug that resulted in the smallest decline in cognitive scores between prescription and the next visit was selected as the treatment of choice. Change of cognitive scores up to 2 years after treatment initiation was compared for model evaluation.

RESULTS

Overall, 1343 patients with MMSE scores were identified for internal validation and 285 [21.22%] took the drug recommended. After 2 years, the reduction of mean [standard deviation] MMSE score in this group was significantly smaller than the remaining 1058 [78.78%] patients (0.60 [0.26] vs 2.80 [0.28]; P = 0.02). In the external validation cohort (N = 1772), 222 [12.53%] patients took the drug recommended and reported a smaller MMSE reduction compared to the 1550 [87.47%] patients who did not (1.01 [0.49] vs 4.23 [0.60]; P = 0.01). A similar performance gap was seen when testing the model on patients prescribed with AChEIs only.

CONCLUSIONS

It was possible to identify the most effective drug for the real-world treatment of cognitive impairment in dementia at an individual patient level. Routine care patients whose prescribed medications were the best fit according to the model had better cognitive performance after 2 years.

Effect of trazodone on cognitive decline in people with dementia: Cohort study using UK routinely collected data

OBJECTIVES

Evidence in mouse models has found that the antidepressant trazodone may be protective against neurodegeneration. We therefore aimed to compare cognitive decline of people with dementia taking trazodone with those taking other antidepressants.

METHODS

Three identical naturalistic cohort studies using UK clinical registers. We included all people with dementia assessed during 2008-16 who were recorded taking trazodone, citalopram or mirtazapine for at least 6 weeks. Linear mixed models examined age, time and sex-adjusted Mini-mental state examination (MMSE) change in people with all-cause dementia taking trazodone compared with those taking citalopram and mirtazapine. In secondary analyses, we examined those with non-vascular dementia; mild dementia; and adjusted results for neuropsychiatric symptoms. We combined results from the three study sites using random-effects meta-analysis.

RESULTS

We included 2,199 people with dementia, including 406 taking trazodone, with mean 2.2 years follow-up. There was no difference in adjusted cognitive decline in people with all-cause or non-vascular dementia taking trazodone, citalopram or mirtazapine in any of the three study sites. When data from the three sites were combined in meta-analysis, we found greater mean MMSE decline in people with all-cause dementia taking trazodone compared to those taking citalopram (0·26 points per successive MMSE measurement, 95% CI 0·03-0·49; p = 0·03). Results in sensitivity analyses were consistent with primary analyses.

CONCLUSIONS

There was no evidence of cognitive benefit from trazodone compared to other antidepressants in people with dementia in three naturalistic cohort studies. Despite preclinical evidence, trazodone should not be advocated for cognition in dementia.

Med7: A transferable clinical natural language processing model for electronic health records

Electronic health record systems are ubiquitous and the majority of patients' data are now being collected electronically in the form of free text. Deep learning has significantly advanced the field of natural language processing and the self-supervised representation learning and the transfer learning have become the methods of choice in particular when the high quality annotated data are limited. Identification of medical concepts and information extraction is a challenging task, yet important ingredient for parsing unstructured data into structured and tabulated format for downstream analytical tasks. In this work we introduced a named-entity recognition (NER) model for clinical natural language processing. The model is trained to recognise seven categories: drug names, route of administration, frequency, dosage, strength, form, duration. The model was first pre-trained on the task of predicting the next word, using a collection of 2 million free-text patients' records from MIMIC-III corpora followed by fine-tuning on the named-entity recognition task. The model achieved a micro-averaged F1 score of 0.957 across all seven categories. Additionally, we evaluated the transferability of the developed model using the data from the Intensive Care Unit in the US to secondary care mental health records (CRIS) in the UK. A direct application of the trained NER model to CRIS data resulted in reduced performance of F1 = 0.762, however after fine-tuning on a small sample from CRIS, the model achieved a reasonable performance of F1 = 0.944. This demonstrated that despite a close similarity between the data sets and the NER tasks, it is essential to fine-tune the target domain data in order to achieve more accurate results. The resulting model and the pre-trained embeddings are available at https://github.com/kormilitzin/med7.

Real-world effectiveness, its predictors and onset of action of cholinesterase inhibitors and memantine in dementia: retrospective health record study

BACKGROUND

The efficacy of acetylcholinesterase inhibitors and memantine in the symptomatic treatment of Alzheimer's disease is well-established. Randomised trials have shown them to be associated with a reduction in the rate of cognitive decline.

AIMS

To investigate the real-world effectiveness of acetylcholinesterase inhibitors and memantine for dementia-causing diseases in the largest UK observational secondary care service data-set to date.

METHOD

We extracted mentions of relevant medications and cognitive testing (Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) scores) from de-identified patient records from two National Health Service (NHS) trusts. The 10-year changes in cognitive performance were modelled using a combination of generalised additive and linear mixed-effects modelling.

RESULTS

The initial decline in MMSE and MoCA scores occurs approximately 2 years before medication is initiated. Medication prescription stabilises cognitive performance for the ensuing 2-5 months. The effect is boosted in more cognitively impaired cases at the point of medication prescription and attenuated in those taking antipsychotics. Importantly, patients who are switched between agents at least once do not experience any beneficial cognitive effect from pharmacological treatment.

CONCLUSIONS

This study presents one of the largest real-world examination of the efficacy of acetylcholinesterase inhibitors and memantine for symptomatic treatment of dementia. We found evidence that 68% of individuals respond to treatment with a period of cognitive stabilisation before continuing their decline at the pre-treatment rate.

Comorbidity between Alzheimer's disease and major depression: a behavioural and transcriptomic characterization study in mice

BACKGROUND

Major depression (MD) is the most prevalent psychiatric disease in the population and is considered a prodromal stage of the Alzheimer's disease (AD). Despite both diseases having a robust genetic component, the common transcriptomic signature remains unknown.

METHODS

We investigated the cognitive and emotional behavioural responses in 3- and 6-month-old APP/PSEN1-Tg mice, before β-amyloid plaques were detected. We studied the genetic and pathway deregulation in the prefrontal cortex, striatum, hippocampus and amygdala of mice at both ages, using transcriptomic and functional data analysis.

RESULTS

We found that depressive-like and anxiety-like behaviours, as well as memory impairments, are already present at 3-month-old APP/PSEN1-Tg mutant mice together with the deregulation of several genes, such as Ciart, Grin3b, Nr1d1 and Mc4r, and other genes including components of the circadian rhythms, electron transport chain and neurotransmission in all brain areas. Extending these results to human data performing GSEA analysis using DisGeNET database, it provides translational support for common deregulated gene sets related to MD and AD.

CONCLUSIONS

The present study sheds light on the shared genetic bases between MD and AD, based on a comprehensive characterization from the behavioural to transcriptomic level. These findings suggest that late MD could be an early manifestation of AD.

ANMerge: A Comprehensive and Accessible Alzheimer's Disease Patient-Level Dataset

BACKGROUND

Accessible datasets are of fundamental importance to the advancement of Alzheimer's disease (AD) research. The AddNeuroMed consortium conducted a longitudinal observational cohort study with the aim to discover AD biomarkers. During this study, a broad selection of data modalities was measured including clinical assessments, magnetic resonance imaging, genotyping, transcriptomic profiling, and blood plasma proteomics. Some of the collected data were shared with third-party researchers. However, this data was incomplete, erroneous, and lacking in interoperability.

OBJECTIVE

To provide the research community with an accessible, multimodal, patient-level AD cohort dataset.

METHODS

We systematically addressed several limitations of the originally shared resources and provided additional unreleased data to enhance the dataset.

RESULTS

In this work, we publish and describe ANMerge, a new version of the AddNeuroMed dataset. ANMerge includes multimodal data from 1,702 study participants and is accessible to the research community via a centralized portal.

CONCLUSION

ANMerge is an information rich patient-level data resource that can serve as a discovery and validation cohort for data-driven AD research, such as, for example, machine learning and artificial intelligence approaches.

Maximizing the use of social and behavioural information from secondary care mental health electronic health records

PURPOSE

The contribution of social and behavioural factors in the development of mental health conditions and treatment effectiveness is widely supported, yet there are weak population level data sources on social and behavioural determinants of mental health. Enriching these data gaps will be crucial to accelerating precision medicine. Some have suggested the broader use of electronic health records (EHR) as a source of non-clinical determinants, although social and behavioural information are not systematically collected metrics in EHRs, internationally.

OBJECTIVE

In this commentary, we highlight the nature and quality of key available structured and unstructured social and behavioural data using a case example of value counts from secondary mental health data available in the UK from the UK Clinical Record Interactive Search (CRIS) database; highlight the methodological challenges in the use of such data; and possible solutions and opportunities involving the use of natural language processing (NLP) of unstructured EHR text.

CONCLUSIONS

Most structured non-clinical data fields within secondary care mental health EHR data have too much missing data for adequate use. The utility of other non-clinical fields reported semi-consistently (e.g., ethnicity and marital status) is entirely dependent on treating them appropriately in analyses, quantifying the many reasons behind missingness in consideration of selection biases. Advancements in NLP offer new opportunities in the exploitation of unstructured text from secondary care EHR data particularly given that clinical notes and attachments are available in large volumes of patients and are more routinely completed by clinicians. Tackling ways to re-use, harmonize, and improve our existing and future secondary care mental health data, leveraging advanced analytics such as NLP is worth the effort in an attempt to fill the data gap on social and behavioural contributors to mental health conditions and will be necessary to fulfill all of the domains needed to inform personalized interventions.

Identifying Predictors of Suicide in Severe Mental Illness: A Feasibility Study of a Clinical Prediction Rule (Oxford Mental Illness and Suicide Tool or OxMIS)

BACKGROUND

Oxford Mental Illness and Suicide tool (OxMIS) is a brief, scalable, freely available, structured risk assessment tool to assess suicide risk in patients with severe mental illness (schizophrenia-spectrum disorders or bipolar disorder). OxMIS requires further external validation, but a lack of large-scale cohorts with relevant variables makes this challenging. Electronic health records provide possible data sources for external validation of risk prediction tools. However, they contain large amounts of information within free-text that is not readily extractable. In this study, we examined the feasibility of identifying suicide predictors needed to validate OxMIS in routinely collected electronic health records.

METHODS

In study 1, we manually reviewed electronic health records of 57 patients with severe mental illness to calculate OxMIS risk scores. In study 2, we examined the feasibility of using natural language processing to scale up this process. We used anonymized free-text documents from the Clinical Record Interactive Search database to train a named entity recognition model, a machine learning technique which recognizes concepts in free-text. The model identified eight concepts relevant for suicide risk assessment: medication (antidepressant/antipsychotic treatment), violence, education, self-harm, benefits receipt, drug/alcohol use disorder, suicide, and psychiatric admission. We assessed model performance in terms of precision (similar to positive predictive value), recall (similar to sensitivity) and F1 statistic (an overall performance measure).

RESULTS

In study 1, we estimated suicide risk for all patients using the OxMIS calculator, giving a range of 12 month risk estimates from 0.1-3.4%. For 13 out of 17 predictors, there was no missing information in electronic health records. For the remaining 4 predictors missingness ranged from 7-26%; to account for these missing variables, it was possible for OxMIS to estimate suicide risk using a range of scores. In study 2, the named entity recognition model had an overall precision of 0.77, recall of 0.90 and F1 score of 0.83. The concept with the best precision and recall was medication (precision 0.84, recall 0.96) and the weakest were suicide (precision 0.37), and drug/alcohol use disorder (recall 0.61).

CONCLUSIONS

It is feasible to estimate suicide risk with the OxMIS tool using predictors identified in routine clinical records. Predictors could be extracted using natural language processing. However, electronic health records differ from other data sources, particularly for family history variables, which creates methodological challenges.

Natural language processing for structuring clinical text data on depression using UK-CRIS

BACKGROUND

Utilisation of routinely collected electronic health records from secondary care offers unprecedented possibilities for medical science research but can also present difficulties. One key issue is that medical information is presented as free-form text and, therefore, requires time commitment from clinicians to manually extract salient information. Natural language processing (NLP) methods can be used to automatically extract clinically relevant information.

OBJECTIVE

Our aim is to use natural language processing (NLP) to capture real-world data on individuals with depression from the Clinical Record Interactive Search (CRIS) clinical text to foster the use of electronic healthcare data in mental health research.

METHODS

We used a combination of methods to extract salient information from electronic health records. First, clinical experts define the information of interest and subsequently build the training and testing corpora for statistical models. Second, we built and fine-tuned the statistical models using active learning procedures.

FINDINGS

Results show a high degree of accuracy in the extraction of drug-related information. Contrastingly, a much lower degree of accuracy is demonstrated in relation to auxiliary variables. In combination with state-of-the-art active learning paradigms, the performance of the model increases considerably.

CONCLUSIONS

This study illustrates the feasibility of using the natural language processing models and proposes a research pipeline to be used for accurately extracting information from electronic health records.

CLINICAL IMPLICATIONS

Real-world, individual patient data are an invaluable source of information, which can be used to better personalise treatment.

A metabolite-based machine learning approach to diagnose Alzheimer-type dementia in blood: Results from the European Medical Information Framework for Alzheimer disease biomarker discovery cohort

Introduction

Machine learning (ML) may harbor the potential to capture the metabolic complexity in Alzheimer Disease (AD). Here we set out to test the performance of metabolites in blood to categorize AD when compared to CSF biomarkers.

Methods

This study analyzed samples from 242 cognitively normal (CN) people and 115 with AD-type dementia utilizing plasma metabolites (n = 883). Deep Learning (DL), Extreme Gradient Boosting (XGBoost) and Random Forest (RF) were used to differentiate AD from CN. These models were internally validated using Nested Cross Validation (NCV).

Results

On the test data, DL produced the AUC of 0.85 (0.80–0.89), XGBoost produced 0.88 (0.86–0.89) and RF produced 0.85 (0.83–0.87). By comparison, CSF measures of amyloid, p-tau and t-tau (together with age and gender) produced with XGBoost the AUC values of 0.78, 0.83 and 0.87, respectively.

Discussion

This study showed that plasma metabolites have the potential to match the AUC of well-established AD CSF biomarkers in a relatively small cohort. Further studies in independent cohorts are needed to validate whether this specific panel of blood metabolites can separate AD from controls, and how specific it is for AD as compared with other neurodegenerative disorders.

Association between Plasma Ceramides and Phosphatidylcholines and Hippocampal Brain Volume in Late Onset Alzheimer's Disease

Lipids such as ceramides and phosphatidylcholines (PC) have been found altered in the plasma of Alzheimer’s disease (AD) patients in a number of discovery studies. For this reason, the levels of 6 ceramides and 3 PCs, with different fatty acid length and saturation levels, were measured in the plasma from 412 participants (AD n = 205, Control n = 207) using mass spectrometry coupled with ultra-performance liquid chromatography. After this, associations with AD status, brain atrophy, and age-related effects were studied. In the plasma of AD participants, cross-sectional analysis revealed elevated levels of three ceramides (Cer16:0 p < 0.01, Cer18:0 p < 0.01, Cer24:1 p < 0.05). In addition, two PCs in AD plasma (PC36:5 p < 0.05, PC38:6 p < 0.05) were found to be depleted compared to the control group, with PC36:5 also associating with hippocampal atrophy (p < 0.01). Age-specific analysis further revealed that levels of Cer16:0, Cer18:0, and Cer20:0 were associated with hippocampal atrophy only in younger participants (age < 75, p < 0.05), while all 3 PCs did so in the older participants (age > 75, p < 0.05). PC36:5 was associated with AD status in the younger group (p < 0.01), while PC38:6 and 40:6 did so in the older group (p < 0.05). In this study, elevated ceramides and depleted PCs were found in the plasma from 205 AD volunteers. Our findings also suggest that dysregulation in PC and ceramide metabolism could be occurring in different stages of AD progression.

Unreliable evidence: 2 sources of uncertainty during perceptual choice

Perceptual decisions often involve integrating evidence from multiple concurrently available sources. Uncertainty arises when the integrated (mean) evidence fails to support one alternative over another. However, evidence heterogeneity (variability) also provokes uncertainty. Here, we asked whether these 2 sources of uncertainty have independent behavioral and neural effects during choice. Human observers undergoing functional neuroimaging judged the average color or shape of a multielement array. The mean and variance of the feature values exerted independent influences on behavior and brain activity. Surprisingly, BOLD signals in the dorsomedial prefrontal cortex (dmPFC) showed polar opposite responses to the 2 sources of uncertainty, with the strongest response to ambiguous tallies of evidence (high mean uncertainty) and to homogenous arrays (low variance uncertainty). These findings present a challenge for models that emphasize the role of the dmPFC in detecting conflict, errors, or surprise. We suggest an alternative explanation, whereby evidence is processed with increased gain near the category boundary.