CHRONOSIG: Digital Triage for Secondary Mental Healthcare using Natural Language Processing - rationale and protocol

Overview

Figure 1 shows an outline user-interface for the clinician-facing CDST. A clinician provides a referral document and (if available) the patient’s existing EHR data (presented to the CDST as a time-ordered sequence of free-text clinical notes). The CDST computes an embedded representation of chronological free-text data (described later). This transforms the patient’s raw free-text data into a space of signatures that exposes patterns of similarity/dis-similarity to the signatures of all patients the CDST encountered during development or ‘training’ of the algorithms (Figure 1, Panel C). A follow-on classifier -- that has been trained to associate signatures with known triage outcomes -- delivers a conditional distribution of continuous triage outcome scores; for example, in Figure 1 (Panel A), scores are represented as a conditional probability distribution of triage outcomes [0.1, 0.3, 0.5, 0.1] for teams W,X,Y and Z respectively. At this stage, a recommendation rule must be applied to these scores. Figure 1 (Panel A) shows one such rule where the maximum value is taken as the recommended triage outcome (i.e. Team Y with score 0.5) and the confidence is shown as inversely proportional to the entropy (Cover, 1999) of this distribution (approximately, if the distribution is heavily ‘peaked’ around a single team, the entropy is low and the confidence high). Note that triage scores need not be mutually exclusive -- in the example shown in Figure 1, Team X might also provide benefit to the patient because the patient’s signature displayed features (for example, co-morbidity) similar to other patients treated by Team X. To assist clinicians, they can then explore features in the patient’s raw data that were salient and driving the classifier’s outputs (Figure 1, Panel B). These components are the foundation for a principled AI-augmented clinical decision making process - rather than the CDST delivering an automated triage decision, the final “decision rule” is a combination of the output of the CDST with the MDT’s clinical expertise.

• Download figure
• Open in new tab

Figure 1:

Prototype functionality and user-interface for CHRONOSIG. A patient referral (with no previous psychiatric history) is triaged by the CDS tool which recommends Team Y (with high confidence; panel A). The clinician can use the document map (Referral Document; panel B) to view specific features that support and explain the suggested triage decision. The middle panel (C) visualises the space of all known patient embeddings (with the current patient highlighted) and a numerical estimate of how similar the current patient is to patients previously referred to Team Y alongside an indication of the likelihood this patient would be accepted given their similarity to other patients accepted by Team Y.

• Download figure
• Open in new tab

Figure 2:

Data Pipeline for CHRONOSIG -- patients’ free-text referral and historical EHR clinical data are used (without a priori human annotation) as input to a language model neural network. The resulting numerical representations of patients’ data (denoted “embeddings” or longitudinal fingerprints) are the input to a moderately-sized downstream supervised neural network classifier that learns associations between treatment teams and the embedding representations.

Patient and Public Co-Production

CHRONOSIG has been developed and funded to provide a patient and public involvement (PPI) work package that runs concurrently with engineering and experimental work. The first PPI deliverable is a stakeholder impact assessment using the FAST (Leslie, 2019) principles (fairness, accountability, sustainability and transparency) in parallel to dataset curation. This output will include PPI (patient and public involvement) definitions of an acceptable referral i.e. what characteristics stakeholders expect should be included or excluded from referrals, for example, features of a patient’s presentation which are thought to negatively bias teams in routine practice.

Data curation is the first milestone in CHRONOSIG and the PPI team will direct the technical team to deliver descriptive analyses relevant to the stakeholder impact assessment. For example, requesting analyses of the representation of gender/diagnosis combinations that, historically, have been poorly served by SMH services as well as providing summaries comparing different Trust’s samples to the Office for National Statistics population census data to query representation of minority groups. This output will be made available as meta-data for the benefit of the wider UK-CRIS network.

The representational scheme employed in CHRONOSIG -- namely, longitudinal signatures of EHR text -- aims to provide high fidelity representations of a given patient’s instance data (equating to e.g. a referral or an historical segment of their EHR history). A disadvantage of this scheme is that although unlikely, there is a theoretical risk of ‘reverse inference’ where signatures contain enough information to either reconstruct elements of the source free-text or identify the patient indirectly. Using word-vector embeddings (Abdalla et al., 2020) demonstrated that up to 68% of patient name/condition pairs could be recovered (i.e. if the model is built from identifiable patient data, associated diagnoses can be gleaned from the representation). For large Transformer language models, again trained on identifiable data, patient name probes generated candidate diagnoses at only the base frequency of each diagnosis in the training data (Lehman et al., 2021) suggesting such models present minimal risk of directly identifying patient/condition pairs. Of course, how NLP models represent broader patient identifiable data (beyond just names) remains an active concern and in CHRONOSIG the PPI group will direct the technical team in stress-testing the signature-generating model to ensure indirect leakage is practically impossible. Additionally, this work will establish that outliers in the space of signatures do not correlate with individual patients or groups of patients with specific protected characteristics (i.e. to prevent undue influence of outliers in the downstream classification model for CDST).

Existing referral and triage processes are not transparent (Chew-Graham et al., 2007), are internal to the SMH team and outcomes and decisions are communicated between the team and the referring party. The National Institute for Health and Care Excellence guidelines on improving patient experiences in mental healthcare (NICE, 2011) recommended studying shared decision making but did not make explicit reference to transparency of the ‘gatekeeping’ role of referral and triage. As a tool for augmented clinical decision making trained on observational data, CHRONOSIG risks recapitulating and entrenching existing practices; given our emphasis on decision making processes we will conduct a patient-led qualitative study during the simulated MDTs. The two primary objectives will be to study existing MDT practices (complementing Chew-Graham et al’s work) as well as observing changes in MDT’s practice when the CDST is present. Using semi-structured observation the patient investigator will capture a) what data are being used by the MDT to triage (i.e. patient’s clinical and demographic features) b) whether triage decisions are made transparent and justified with clinical reasoning and c) when there is disagreement among team members, what drives this e.g. patient factors such as diagnostic or risk history alongside MDT factors such as task and process conflict (Jehn & Mannix, 2001).

Data Curation

The UK-CRIS network (https://crisnetwork.co/uk-cris-programme) enables the curation of datasets of de-identified patient EHR records with well-established operating procedures for secure data curation/storage, ethics, information governance and cyber-security. CHRONOSIG will curate a sample from four demographically different NHS Trusts in the UK-CRIS network for model development and validation with a completely held-out and vaulted sample for model testing.

Patients will be sampled on the basis of types of community teams they were referred to and treated by. The curated dataset will contain unstructured free text records (so-called “progress notes” which capture referral information, details of previous clinical assessments/treatment and administrative contacts with the patient) alongside structured data representing CMHTs the patients were treated by (i.e. a proxy for triage outcomes).

Each patient instance is the unstructured text data from the first time-stamped EHR entry through to the point of a given referral to a secondary care service. A single patient having multiple contacts/treatment episodes with SMH services will be represented by many instances, each one containing the cumulative clinical history from the start of the patient’s EHR record up to the timestamp of a referral or episode of care (reflecting the data a team would have access to when triaging a given referral). When a patient was admitted to hospital (for example, in an emergency) these episodes are not counted as a referral because these are triaged and managed differently from referrals to community services (which make up the majority of secondary mental health care and are the focus of this project). However, the consequent unstructured inpatient clinical data will remain in any future instances if the patient was subsequently referred to community care because inpatient progress data would be available to a triaging MDT.

Data Bias and Algorithmic Fairness

Similar to data collected by other public sector bodies (ONS, 2018), healthcare records rarely have reliable data recording for ethnicity, gender identity, sexual orientation, relationship status and culture identity. Internationally, people from ethnic groups that are minorities in their domiciled countries are more likely to be admitted for involuntary psychiatric care (Barnett et al., 2019). Surveys conducted with LGBTQ+ communities show susceptibility to specific mental health problems as well as identifying prejudice, stigmatisation and discrimination from healthcare organisations as being reasons people do not seek input for mental health problems (Stonewall, 2018). Together, these factors predict lower representation in our curated dataset and by construction, any statistical model of these properties will be biased. In CHRONOSIG, we will first provide a descriptive audit of available data in UK-CRIS to describe recording for under-represented people compared to population and census data. This will provide meta-data to the wider community making use of UK-CRIS data.

There is no single technical solution to the problem of biased source data. For the derived signature representations, we will use dimensionality reduction methods to visualise networks of signatures with patients explicitly labelled for possessing protected characteristics or being members of under-represented groups -- this enables us to determine if the signature representations can differentially ‘expose’ patients that may be detrimental in the downstream classification (triage outcome) pipeline.

Inferring unobserved and protected characteristic data is demonstrably problematic (Tomasev et al., 2021) -- for example, the risk of unintentional disclosure or an incorrect assumption of predicted gender identity during a triage activity using the CDST. Given we cannot ‘correct’ for under-representation in data sources we will apply principles of distributive justice for algorithmic fairness (Rajkomar et al., 2018) specifically:

• we will not defer triage decisions to an algorithm -- the CDST delivers transparent recommendations whose fairness (and statistical performance) are only evaluated in the context of MDT decision processes (which themselves, may display unfair biases)

• by examining clinical decision making using simulated triage MDTs (described below) we will employ fairness elicitation (Jung et al., 2020) as a partial solution to deploying constraints on CDST outputs to promote individually equitable triage for all patients

Natural Language Processing for Longitudinal Patient Representation

Various NLP methods have been developed and applied to represent and extract clinically meaningful information from chronologically ordered patients’ free-text records (Dalianis, 2018). The common approach was to train an information extraction model (Kormilitzin et al., 2021; Wang et al., 2018) to recognise a predefined set of clinical concepts of interest (e.g. medications, symptoms, health conditions). Once recognised, the concepts were organised as a temporal knowledge graph (Kormilitzin et al., 2020; Leetaru & Schrodt, 2013) that can be queried (Senior et al., 2020).

However, the development of contextual word representations and Transformer-based language models, such as BERT (Devlin et al., 2019) have led to numerous “end-to-end” applications where embedded representations of clinical text inputs are used directly for downstream analytical tasks, such as disease progression modelling and clinical decision support (Bai et al., 2018).

More recently, very large language models (Brown et al., 2020) and task-specific prompting, demonstrated a promising approach to maximise rich information contained in voluminous textual data. However, BERT-like models suffer from the input text length limitation of 512 tokens, whereas the average document length of one year-worth patient’s notes may contain on average 11,000 tokens. Several mitigation strategies were introduced to learn from long texts, including a hierarchical chunking of long texts (Zhang et al., 2020) and sparse attention mechanisms (Zaheer et al., 2020)

We will exploit a recently proposed mathematical methodology of the low-rank tensor approximations (Toth et al., 2020a) to augment the sparse attention mechanism. The new attention mechanism will leverage rich information contained in long chronological EHR to represent patients’ trajectories and ensure the training is computationally efficient. The low-rank tensor approximation for sequential data has already showed promising results (Toth et al., 2020b) on standard benchmark datasets and will allow to capture the long-range dependencies in chronological EHR whereby tokens are separated further apart to learn the longitudinal signature of patients’ health and interventions. The pre-trained language models adapted for long texts, will encode a chronological collection of input free-text patients’ records into a high-dimensional feature vector, representing the longitudinal patients’ signature for downstream predictive tasks.

Simulated MDTs for Evaluating Augmented Decision Making

Our proposal is that clinical decision making should be augmented -- not replaced -- by CDSTs because historical emphasis on automated decision making has under-delivered (Longoni et al., 2019; Nagendran et al., 2020). Clinical decision making requires at the very minimum, knowledge of the risks of triage decisions i.e. a function of the probability of -- and the costs associated with -- each potential outcome. For example, if an MDT rejects a referral, they will be implicitly estimating the risks for the patient weighed against those of accepting an inappropriate referral; while one could easily estimate a resource cost of accepting the referral, the cost of catastrophic outcomes -- such as the death of a patient -- cannot be meaningfully captured in the same “currency” (Vickers & Elkin, 2006). Additionally, even in well-calibrated predictive models, assumptions implicit in algorithmic decision rules (i.e. that health resource cost is an appropriate proxy for health needs) can lead to inequitable outcomes across different racial groups (Obermeyer et al., 2019).

In CHRONOSIG, summary performance metrics (e.g. triage outcome accuracy and error rates) will only be used to evidence that models are learning from the available data. The performance of the tool will be evaluated in controlled experiments where triage decisions from simulated MDTs with and without the CDST are compared. Simulated MDTs will be composed of mental health professionals representative of clinical teams who make triage decisions in secondary mental health care. In a single experiment, patient instances from the vaulted testing sample (i.e. with known triage outcomes but not used during model development) will be presented to both MDTs for triage. The primary outcome will be the between-MDT difference in time spent on each patient’s triage. Secondary outcomes will include analyses of a) within-MDT agreement (homogeneity of the team’s decision making) between-MDT agreement (to establish if triage decisions are automation-biased by having the CDST) and c) identifying pathological triage cases where e.g. the CDST recommendations are consistently at odds with the MDTs decisions. A further qualitative study of the MDT process was described above.

Patient Safety. Governance and Regulatory Factors

For this development project, it would be premature and a risk to patient safety to prospectively evaluate effectiveness with ‘live’ triage cases for clinical use. Instead, validation, testing and initial efficacy data for the CDST will be collected using simulated MDTs with data from historical patients in the UK-CRIS database. The project benefits from the federated UK-CRIS network and inherits the infrastructure and a mature information governance framework. Patient-level EHR data in the UK-CRIS database is pseudonymised and uniquely identifying data (e.g. NHS numbers, dates of birth and names) are masked prior to data being warehoused for research use. Access to UK-CRIS data is strictly controlled to researchers approved by each contributing NHS Trust and there are patient opt-out mechanisms in place. From the perspective of deploying the CDST tool, no source EHR data is required (or stored) for deployment and as described above, risks of reverse inference for patient identifiable data (Abdalla et al., 2020; Lehman et al., 2021) will be investigated thoroughly.

Deployment at Scale

Almost all modern neural network-based predictive models require substantial computing resources in excess of the typical IT capacity available at a clinician’s desk - a problem well recognised in the NHS (BMA, 2019) but receiving less emphasis in forward-looking state-of-the-art surveys (Joshi & Morley, 2019). Therefore, if this project succeeds and the initial efficacy data is positive, we will migrate the CDST to a cloud-service to facilitate a full parallel-arm prospective controlled trial to include health economic impact. Moving to cloud-based services also mitigates the carbon footprint associated with training language models (Bender et al. 2021) that is essential to the development of responsible AI technology and consistent with the NHS’ ambitions for minimising climate impact.