Reply to: dopaminergic connectivity and POCD: balancing statistical robustness with pathological sensitivity

Editor, We are grateful that our manuscript has attracted such strong interest and appreciate Dr Jiang's thoughtful critique.1 The dopaminergic hypothesis of postoperative neurocognitive disorders has long been postulated; however, the interactions neurotransmitters, anaesthetic toxicity, surgical trauma and brain function remain unclear. As our colleague points out, future research needs consider advanced modelling methods, availability of digital clinical routine data and novel monitoring methods, such as wearable devices that collect data with high temporal resolution. Dr Jiang suggests that partial least squares (PLS) regression provides a more powerful statistical approach to our research question than principal component analysis (PCA). While PCA models variance within a set of variables, PLS models covariance between a set of independent variables (e.g., a functional connectivity dataset) and a dependent variable (e.g., POCD). However, PLS is not a panacea for lack of strong hypotheses in big data analysis, and by default, it does not discriminate between signal and noise. Hence, we would like to question Dr Jiang's generalised statement that PLS provides “biologically more interpretable predictors”. Indeed, PLS models suffer from the same tendency towards overfitting to noise as LASSO or Ridge regression in the data-driven analysis of large datasets. Building PLS models requires rigorous training, testing and validation in independent samples to select hyperparameters and assess model validity and overfitting.2 Therefore, the sensitivity of PLS regression depends strongly on the sample size and event rate, and the same concerns against Ridge or LASSO regression apply to PLS regression. As long as larger datasets are not available, other approaches to address this research question may be discussed. Future analysis may not assess postoperative cognitive outcome as a binary variable but as a continuously scaled outcome (e.g., global cognitive component ‘g’) to address the issue of low event rates. While this may allow more complex analyses, the relevance of subtle changes on a continuous cognitive outcome may be a matter of debate: In the recent recommendations for the nomenclature of postoperative neurocognitive disorders (PNCD), major PNCD requires cognitive decline and functional impairment to ascertain that the assessed outcome impacts the patients’ lives.3 Dr Jiang further outlines that precise data on blood pressure fluctuations during surgery and exposure to anaesthetics are mandatory to fully understand dopaminergic dysfunction in postoperative neurocognitive disorders. With the advent of healthcare digitalisation, we look forward to high-resolution routine clinical data becoming available to complement research data. We would also like to emphasise that these data should be recorded throughout the postoperative period with wearable devices. The assessment should also be extended to more elaborate measures of vegetative outflow, such as heart rate variability,4 and intra-operative burst suppression as a measure of cerebral hypometabolism.5 In addition, functional magnetic resonance imaging (fRMI) of the midbrain and brain stem is challenged by respiratory motion, and CO2-dependent changes in blood flow.6 Ideally, haemodynamic monitoring should be extended to fMRI acquisition to address any bias between fMRI preprocessing and haemodynamic stability in the peri-operative period. We further agree with our colleague that survivor bias may have impacted our results. While Dr Jiang points out that patients with cognitive decline probably dropped out of fMRI more often, we would like to add that patients with cognitive decline may be assumed to be lost to follow-up in general, including refusal or early discontinuation of cognitive testing and unavailability for visits due to enduring hospitalisation or care dependence. To elaborate on this, we would like to point out that in our cohort, postoperative delirium (POD), the major risk factor for postoperative cognitive decline, was not associated with POCD. In a recent analysis of the same study, we found that POD led to increased loss to follow-up and mortality three months after surgery.7 Hence, it may be reasonable to assume that inability to attend MRI was not the major reason for patients who experienced POD and POCD to complete follow-up. Therefore, we would like to challenge the suggestion of data imputation as a solution to this problem. Eventually, we need structured clinical pathways for surgical patients to be implemented in routine anaesthetic care. These should include the postoperative assessment of cognitive function and transfer to specialist care (e.g., memory clinics) to incorporate data on long-term cognitive trajectories into our research.8 Implementation would also grant access to follow-up MRI data collected as part of routine diagnostic procedures. We welcome the suggestion of structural equation modelling and especially appreciate the underlying acknowledgement of postoperative neurocognitive disorders as a multifactorial, mediated processes, including at least two levels: Clinical factors – for example, anaesthetic toxicity, surgical trauma with haemodynamic instability and postoperative stress Molecular mechanisms – for example, network and neurotransmitter dysfunction, metabolic and neuroendocrine disruption including disturbances of the circadian rhythm, altered immune system function, senescence and impaired molecular clearance mechanisms, such as autophagy While the former may only be partially modifiable in acute care settings, the latter are potential targets for molecular interventions. We agree with Dr Jiang regarding the potential of omics to unravel the molecular mechanisms underlying clinical factors.7 However, it will be challenging to integrate the suggested abstract concepts of pre-operative reserve, and intra-operative sensitivity – or vulnerability – into these models. In addition, combining molecular and clinical data exponentially increases the complexity of the model. Precise knowledge of causality and pathways is required before creating a confirmatory model. Unfortunately, the commentary by Dr Jiang is missing a roadmap between the choice of the adequate regression model for our exploratory analyses, and the confirmation of complex, molecular and mechanistic models. In summary, we believe that we share a very similar perspective on the challenges in studying postoperative neurocognitive disorders. The suggestions of Dr Jiang will help develop a roadmap for future research, as currently available datasets may not fully provide the required sample size, experimental paradigms, data density and temporal resolution. Therefore, a final answer to the question of which factors and molecular mechanisms evoke postoperative neurocognitive dysfunction is still on the road ahead.