Page 24 - Shimadzu Journal vol.9 Issue1

Clinical Research

This could prove crucial when expanding search of sample qual-
ity biomarkers into other pre-analytical factors such as freeze thaw
cycles, hemolysis, micro-clotting, long-term storage or into other
sample types such as serum or urine.
Only one feature stemmed from the pos mode (pos 974.8), all
other features were negative ions. The neg features consistently
showed larger differences with time_delay. The neg mode alone
would suffice to predict the time_delay. Nevertheless, we would
not yet recommend to abandon the pos mode in future studies.
The risk is too high that highly predictive features for other sample
quality factors (e.g. freeze thaw cycles, hemolysis) could be over-
looked because the combination of both modes notably expands
the covered chemical space. However, in routine application obvi-
ously the reduction to one mode would be very beneficial because
Figure 3. Plasma preparation delay is detectable from selected PESI features measurement time would be halved, doubling throughput.
with high specificity. The best performing feature neg 88.99 stems probably from
A: Both PCA and OPLS-DA scores plot show a clear and highly signifi-
cant difference for time_delay 0 h vs. 3 h, when based the 18 most important lactate. Lactate is also well known to increase with time as an end
features (LOG). B: All five applied machine learning algorithms delivered product of erythrocyte driven glycolysis . The neg 88.99 perfor-
8,10
excellent predictions of time_delay (AUV>0.95) with no false negatives and mance alone would suffice for very good prediction of a 3 h time
very similar ROC curves.
delay. Prediction were as good without the neg 88.99, showing that
the suggested features form a pattern robust against single feature
very beneficial for future routine applications. Nevertheless, we failures. Robustness against single feature failures is important for
would not yet recommend to abandon the pos mode. The combi- routine high-throughput applications reducing the need for re-
nation of both modes notably expands the covered chemical space. peated measurements. Additionally, medical conditions possibly

Figure 4. Heatmaps and plots of the 18 selected features in LOG data.
A: Heatmaps with hierarchical clustering underline the clear difference induced by time_delay in the 18 most important features. There
are no systematic differences between study subgroups or genders. B: Scatter plots showing that from the 18 features, most (16) increased
after the 3 h time_delay while only two feature levels decreased.

23
Shimadzu Journal vol.9 Issue1 23