Can ECT and rTMS Finally Help Us Trust in Precision Psychiatry?

Abstract

As a medical discipline, psychiatry has long grappled with the concept of trust. While there are many reasons for this, it remains a contemporary issue for two of our most crucial treatments for treatment-resistant conditions: electroconvulsive therapy (ECT) and repetitive transcranial magnetic stimulation (rTMS).

This lack of trust has translated into divergent and polarised patient and media narratives for ECT despite extensive evidence of its highly effective treatment [1, 2]. With increased understanding of mechanisms and sophistication of treatment, the clinical procedure of prescribing ECT is also becoming more challenging. As such, clinicians require growing trust in their ability to determine and, to some degree, predict which type of ECT (determined by lead placement, pulse width and stimulus dose in relation to threshold) is likely to lead to remission with additional consideration of obtaining a favourable side-effect profile [3]. For the practising ECT clinician, complexity increases where the specific use of anaesthetic and augmenting agents, selection of titration protocols and procedural timings need to be considered [4]. This is particularly important since the elements of ECT practice that require rating of features are often more impacted by the practitioner's experience level [5].

As we approach 40 years since its inception, the issue of trust in rTMS is less linked to stigma and external factors where it is easier to directly demonstrate modulation of neuronal activity, which the patient can observe. However, trust in selecting optimal treatment parameters remains a subject of intense research after all this time. Although the choices will sound similar (target site, pulses and number of sessions), the fundamental parameters considered, in addition to potential target brain structures, remain the same [6]. The issue is that for both life-saving treatments, there can be ambiguity around whether new patients should start ECT, rTMS or an alternative treatment like ketamine. Then, if they do, a number of optimal treatment parameters need to be decided by the clinician with limited ability to personalise this to the patient.

Plenty of lofty promises have been made about the potential of digital psychiatry. However, one of the biggest is to use machine-learning algorithms to step beyond the capabilities of traditional statistics and reveal connections that have previously not been apparent to us [7]. Despite the promise, machine learning has been criticised for not readily demonstrating to clinicians how individual factors influence the output, leading to a lack of transparency, understanding and mistrust from clinicians using them. This is particularly difficult when models are later demonstrated to have a negative impact, and the reasons cannot be thoroughly dissected. Furthermore, most precision psychiatry studies have remained pilot projects and have never made it into clinical practice [8].

In this edition, two articles address these issues around trust, which will hopefully progress the leap into clinical utility for precision psychiatry for ECT and rTMS. The first paper by Blanken repurposes data from two previous RCTs to perform a network analysis not only to predict remission from depression using ECT through baseline symptoms but also to provide an assessment of the impact of these individual symptoms (referred to as nodes). This study also highlights the importance of considering the passing of time between data points in machine-learning analysis. In this way, they are not only able to identify suicidal ideations, retardation and hypochondriasis as predictive factors but also help build trust by quantifying their impact on the model [9].

Oostra addresses the uncertainty around treatment parameters through a large meta-regression and meta-analysis study comparing the relative impact of more treatment sessions compared with more pulses for high-frequent or low-frequent rTMS of the left or right dorsolateral prefrontal cortex (dlPFC) in treatment-resistant depression. They identify the largest mean difference in the groups receiving 1200–1500 and 360–450 high-frequency and low-frequency pulses per session, respectively. Adding increased trust in this number of pulses, the authors shift the focus to session numbers [10].

Considering the building momentum, how do we ensure the potential of precision ECT and rTMS can be fulfilled? Machine learning algorithms rely on two things. The first is a large, high-quality dataset, and the second is useful clinical questions. We have already covered above that neurostimulation has important clinical questions that still need to be explored. Blanken has already demonstrated the utility of reusing two high-quality RCTs, and ECT is no stranger to large consortiums and datasets [11, 12]. Though these often focus on different things, data on imaging, EEG, clinical and demographic data, blood results, genetics and other physiological results at various time points could all provide relevance to treatment factors. For trust to increase, ECT and rTMS need to utilise their multimodal nature to combine their vast sources of data with explainable machine learning (often referred to as ‘explainable Ai’ or ‘xAi’) as demonstrated in the network approach by Blanken, where each node's impact can be explained [13].

Maturity is also becoming apparent with the standardisation in the reporting of machine learning studies through tools like the Transparent Reporting of multivariable prediction models for Individual Prognosis Or Diagnosis with Artificial Intelligence (TRIPOD + AI) [14]. However, studies so far generally look at prediction performance alone or compare this to the performance of expert clinicians (known as expert in the loop). For clinicians to start trusting the machine, new approaches need to be considered regarding how the two can work together to boost the performance of predictions.

While this creates exciting possibilities, it also raises some interesting questions for us to ponder. How do we create the right environments for psychiatrists and machine learning experts to work together? Expanding on this, how much should we expect the average psychiatrist to understand about digital psychiatry to be able to trust the intervention delivered? Regardless of the final approach, we must accomplish it in a way that does not add any patient or public mistrust in psychiatry.

The paper was conceptualized, written, and approved by R.M.L.

The author declares no conflicts of interest.