5 Conclusion and future research
With this PhD I have tackled challenges in fluid responsiveness prediction during surgery. This was done through three different types of papers: Paper 1 critiques a specific research method, Paper 2 introduces a statistical method and presents novel uses that may be relevant for fluid responsiveness prediction, and Paper 3 is a clinical experiment.
Paper 1 unravels a methodological problem associated with a specific method for predicting fluid responsiveness: the MFC. The bias that occurs in these studies due to shared measurement error, is a common pitfall in studies of change in any measurement. While the problem is especially clear in the MFC studies, this statistical issue should be keep in mind when designing any study where change is an outcome of interest.
Paper 2 introduces GAMs as a tool for analysing medical time series and waveforms. Two motivated examples are presented. The first is relatively simple and has a direct clinical use: by modelling a PP time series with a GAM, we can derive PPV in a robust manner that seems to work even with few beats per ventilation (low HR/RR). In the second example, we demonstrate that a GAM can be used to decompose a CVP waveform into cardiac and ventilatory components.
Paper 3 presents a clinical study of the effects of \(V_T\) and RR on PPV. The results indicate that using a GAM derived PPV may enable fluid responsiveness prediction despite low HR/RR, and that it may be valuable to correct PPV for \(V_T\).
5.1 Future research
Future studies of the MFC should take great care to minimise the problems presented in Paper 1. I am aware of only one MFC study published after the publication of Paper 1 [99]. This study used the same problematic design as previous MFC studies—reportedly to ease comparison with previous MFC studies. They do, however, acknowledge that the design might be problematic.
Also, it may be possible to do an unbiased reanalysis of data from the existing MFC studies with only three SV measurements. I do not have a complete recipe for this—nor the data—but a simple regression model of absolute SV measurements, rather than changes (\(\Delta SV\)), is probably a good starting point. Example: \[ SV_{500} = \beta_1 \cdot SV_{100} + \beta_2 \cdot SV_{baseline}. \] This is essentially a second-order autoregressive model. The part that is still unclear to me, is how to turn the fit of this model into a simple clinical tool and how to report its accuracy.
The GAM-method for estimating PPV, presented in Paper 2, should be evaluated as a predictor of fluid responsiveness in future studies. To include patients with a higher average fluid response, the study protocol should be shorter than what was used in Paper 3. A fast protocol, with e.g. a two minute baseline measurement of both SV and PPV, immediately followed by a fluid challenge, would allow evaluation in a more hemodynamically unstable state. This setup could be run in an ICU, with ARDS patients on lung-protective ventilation, to assess the performance of GAM-derived PPV in a setting with low HR/RR.
The data presented in Paper 3 contains four \(V_T\) challenges for each subject (a \(V_T\) challenge analyses the effect on PPV of increasing \(V_T\) from 6 to 8 ml kg-1). There is currently a medical Master’s project using this data to investigate the effect of RR on the response to the \(V_T\) challenge.
In Paper 2, we also describe how a GAM can be used to decompose a waveform. A proposed use case of this technique is that the \(x'\) descend of the CVP may represent the right ventricle’s contractility, and that a high respiratory variation in this feature could indicate that the right ventricle is sensitive to changes in afterload. It would be interesting to investigate the association between the slope of the \(x'\) descend and TAPSE (ultrasound derived) before and after an increase in PEEP, and whether the respiratory variation in the \(x'\) descend could predict the hemodynamic effect of the PEEP increase.
5.2 General challenges in intraoperative fluid responsiveness prediction
The accuracy of any fluid responsiveness study hinges on the accuracy of the CO or SV measurement. These measurements are generally not very accurate (as described in Section 3.4.3), and uncertainty of the outcome sets a hard limit to how accurate a predictor can be. It would be highly valuable to do rigorous investigations of the agreement between different methods for measuring CO, in a setting where we can trust the gold standard, e.g. using Fick’s principle for CO estimation in a very stable setting or using magnetic resonance velocity mapping [100].
In fluid responsiveness studies, the outcome is generally dichotomised into fluid responders (e.g. \(\Delta SV > 15 \%\)) and non-responders. This is probably done to make the results simpler to interpret, but the choice of the threshold is somewhat arbitrary, and dichotomisation decreases the power of the analysis [101]. It is possible to have the power of the continuous analysis, but report the prediction accuracy dichotomously, by, e.g., reporting the PPV value where 80% of the prediction interval for \(\Delta SV\) is above 15%.2 It would also be interesting to start analysing fluid responsiveness prediction studies using absolute measurements of flow rather than changes. Changes are notoriously difficult to work with [102,103], and relative changes (e.g. a 15% increase from baseline) can be especially problematic [104]. If we use a regression model to predict the body weight indexed SV (SVI) after a fluid challenge using PPV and the baseline SVI as predictors, we can still evaluate the predictive ability of PPV, and we can investigate whether the baseline SVI level modifies this predictive ability.
It is important to be aware that the predictive performance, e.g. AUROC, reported from a fluid responsiveness prediction study is not simply a measure of the test, but a measure of the test in that setting [105]. In some fluid responsiveness studies there is a wide, or even bimodal, distribution of fluid responses, with changes in SV or CO ranging from \(-10\%\) to \(50\%\), and the predictor of choice (e.g. PPV) accurately distinguishes fluid responders from non-responders [53,54,106,107]. On the other hand, in studies where most subjects have a fluid response close to the threshold (e.g. a \(\Delta SV\) of 15%), responders and non-responders are more similar, and the AUROC will usually be lower. The authors may then conclude that the investigated predictor did not work, even though the results from the two types of studies may be entirely compatible, with one study essentially investigating the most difficult subset of the other study. This would be clearer, if we compared parameters from continuous regression analyses across studies—rather than comparing dichotomous classification characteristics like AUROC, sensitivity and specificity.