Computational protocol: A data-driven approach for evaluating multi-modal therapy in traumatic brain injury

Similar protocols

Protocol publication

[…] Spatial learning and memory was assessed with the Morris water maze test. The rats’ ability to locate a submerged platform (10 cm diameter, 2.5 cm below the water surface) was measured in a circular pool (180 cm diameter, 60 cm depth). Rats were trained to locate a visible platform during the first day (indicated by a flag hanging above the platform), followed by the hidden-platform training for additional 4–7 days. During both the visible and hidden platform training, 6 daily trials, counter-balanced for distance and drop locations, were performed. Latency and the distance to locate the platform and swim velocity were measured via a video tracking system (Ethovision, Noldus Information Technology, Sterling, VA). Data from 4 days of hidden platform training were included for PCA analysis. One day after the completion of the hidden platform training, rats were subjected to the last trial in which the hidden platform was removed to assess retention memory (i.e., the probe trial). An annulus with a diameter of 20 cm was defined concentric with the previous location of the hidden platform. In the probe trial, the time spent, the number of crossings and the latency of the first crossing were measured for the annulus, the target quadrant and the platform area. In all sessions (i.e., visible platform, hidden platform and probe trial) thigmotaxic behavior, swimming within 20 cm of the edge of the pool, was determined. [...] Statistical analyses were performed using SPSS Statistics 23 (IBM). We performed three distinct analyses. First, we performed traditional univariate analyses of all outcome measures of interest. We tested the main effect and interactions of drug (i.e., minocycline and LM11A-31) and physical therapy intervention (i.e., no physical therapy, physical therapy alone and physical therapy in combination with botox). For linear variables a LMM was applied. Count data were assessed by a generalized linear model (Poisson probability distribution, log link function). To test how many of the univariate effects would survive multiple-testing correction we used Bonferroni correction and the Benjamini-Hochberg method. Second, to remain sensitive to multimodal changes induced by the combinatorial interventions and to use the full set of outcome measures, an unbiased data-driven multivariate approach was applied. The data-driven analysis was performed without prior assumptions; therefore a primary outcome measure was not defined. An independent analyst blinded to the experimental condition performed all data-driven analyses. Third, after multidimensional cross-validation of outcome patterns experimental conditions were decoded for explicit hypothesis testing using a LMM on PCA derived scores (PCA-LMM) (). The interaction effects (i.e., synergistic and counteracting) of different drug interventions and physical therapies were all assessed using this linked analytical workflow.To integrate data across outcome measures with diverse scales we used a non-linear PCA (NL-PCA). A tutorial on how to apply NL-PCA has been provided by Linting et al. and the SPSS syntax for the current NL-PCA is available as . NL-PCA is suitable for variables of mixed measurement levels (nominal, ordinal and numeric). In NL-PCA variables are assigned numerical values through a process called optimal scaling transformation. Optimal scaling assigns quantitative values to categorical variables in an optimal way, meaning maximizing the variance of the predefined number of PC (i.e., dimensions). The NL-PCA dimension reduction and stability testing workflow is shown in . NL-PCA was initially applied using a 6-dimensional solution. The final dimensionality of the PCA (i.e., number of principal components) was defined based on the following criteria: (1) Kaiser rule, retaining principal components with eigenvalues >1; (2) Cattell rule, retaining principal components above the elbow in the scree plot; (3) PC over-determination, retaining components with at least four loadings above 0.6. Based on these criteria, the final number of PC were defined to maximize the variance accounted for and at the same time reducing the number of components (). In NL-PCA the selected dimensionality affects analysis results. Therefore, after determining the number of PC to retain the PCA was re-run with the final dimensionality.Stability of PCA solutions was further assessed by internal and external cross-validation (). To determine the internal stability of the NL-PCA solution, we performed nonparametric balanced bootstrapping procedure using 2000 iterations and Procrustes rotation. The non-bootstrapped PCA solutions assessed for cross-validation against the bootstrapped PCA solution by using pattern matching statistics: root mean square difference in PC loading patterns, the coefficient of congruence, the Pearson product moment correlation coefficient, and the Cattell salient variable similarity index. Convergence of these metrics indicates consensus of highly-reproducible PC patterns. Though there exist no clear threshold values for the root mean square and the coefficient of congruence, it has been suggested that root mean square values approaching 0.0 and coefficient of congruence approaching 1.0 indicate an appropriate fit of both the magnitude and the sign of the loading patterns. In the current study, PC loading pattern matching was set at p < 0.05 for both Pearson r and the salient variable similarity index (s). For the salient variable similarity index, we used the conservative cutoff of |.4| for assessing salient loading. For the external cross-validation the dataset was split up into its separate experiments (i.e., Double-combo Mino, Double-combo LM11A-31 and Triple-combo). PCA was then run on each experiment separately and the resulting PCA solutions were compared using the same pattern matching statistics as for the internal cross-validation.All NL-PCA were performed in and unsupervised manner by an analyst who was blind to experimental condition. Only after cross-validation were experimental conditions decoded for explicit hypothesis testing. Hypothesis testing LMM was used on the stable PC (based on internal and external cross-validation) to test the effect of drug intervention and physical therapy intervention (i.e., no physical therapy, physical therapy alone and physical therapy in combination with botox) on principal component scores, thereby assessing the full set of functional outcome measures and brain markers simultaneously (). Significant effects were followed by Tukey’s post hoc test of group means. Statistical significance for all analysis was set at α = 0.05. […]

Pipeline specifications

Software tools EthoVision, SPSS
Applications Miscellaneous, Macroscope & basic digital camera imaging
Organisms Rattus norvegicus, Homo sapiens
Diseases Brain Injuries
Chemicals Minocycline