Detecting the Hidden Properties of Immunological Data and Predicting the Mortality Risks of Infectious Syndromes

Abstract

Background: To extract more information, the properties of infectious disease data, including hidden relationships, could be considered. Here, blood leukocyte data were explored to elucidate whether hidden information, if uncovered, could forecast mortality.

Methods: Three sets of individuals (n = 132) were investigated, from whom blood leukocyte profiles and microbial tests were conducted (i) cross-sectional analyses performed at admission (before bacteriological tests were completed) from two groups of hospital patients, randomly selected at different time periods, who met septic criteria [confirmed infection and at least three systemic inflammatory response syndrome (SIRS) criteria] but lacked chronic conditions (study I, n = 36; and study II, n = 69); (ii) a similar group, tested over 3 days (n = 7); and (iii) non-infected, SIRS-negative individuals, tested once (n = 20). The data were analyzed by (i) a method that creates complex data combinations, which, based on graphic patterns, partitions the data into subsets and (ii) an approach that does not partition the data. Admission data from SIRS+/infection+ patients were related to 30-day, in-hospital mortality.

Results: The non-partitioning approach was not informative: in both study I and study II, the leukocyte data intervals of non-survivors and survivors overlapped. In contrast, the combinatorial method distinguished two subsets that, later, showed twofold (or larger) differences in mortality. While the two subsets did not differ in gender, age, microbial species, or antimicrobial resistance, they revealed different immune profiles. Non-infected, SIRS-negative individuals did not express the high-mortality profile. Longitudinal data from septic patients displayed the pattern associated with the highest mortality within the first 24 h post-admission. Suggesting inflammation coexisted with immunosuppression, one high-mortality sub-subset displayed high neutrophil/lymphocyte ratio values and low lymphocyte percents. A second high-mortality subset showed monocyte-mediated deficiencies. Numerous within- and between-subset comparisons revealed statistically significantly different immune profiles.

Conclusion: While the analysis of non-partitioned data can result in information loss, complex (combinatorial) data structures can uncover hidden patterns, which guide data partitioning into subsets that differ in mortality rates and immune profiles. Such information can facilitate diagnostics, monitoring of disease dynamics, and evaluation of subset-specific, patient-specific therapies.

Keywords: complexity; immunomicrobial interactions; immunosuppression; pattern recognition; sepsis; visual.

Figure 1

Non-partitioning-based (non-combinatorial) analysis of leukocyte and biomarker data. The outcomes reported within 30 days after SIRS+, infected individuals were admitted were not differentiated by input (blood leukocyte) data collected at admission. Both counts (A,B), percentages (C,D), and ratios of leukocyte cell types (E,F) as well as CRP concentrations (G,H) did not distinguish survivors from non-survivors: data overlapping was observed in both studies (blue boxes).

Figure 2

Elimination of non-informative patterns. The combinatorial and three-dimensional (3D) method was not always informative: many plots did not show distinct patterns (A,B), even when a single (one data point-wide) line of observations was detected (C,D).

Figure 3

Three-dimensional (combinatorial and partitioning-oriented) analysis of dimensionless indicators derived from leukocyte data (structure I). The large number of combinations the alternative method can generate increased the likelihood of finding some informative plots. Both study I (A) and II (B) displayed a single (one data point-wide) line of observations, which consisted of two segments perpendicular to one another. Such graphic data structure improves detection because data points can only occur along the line.

Figure 4

Mortality rates of perpendicular data segments. Both study I (A) and II (B) displayed mortality rates at least twice higher in the subset located on the right side of the plot than in the left subset (66.6% in both studies vs. 33.3 or 23.1%, in study I or II, respectively). Such differences approached or achieved statistical significance [P = 0.057 (study I) or P ≤ 0.01 (study II), Chi-square test]. When the same data structure was utilized to analyze 20 non-infected, SIRS-negative individuals, the high-mortality subset was not observed (C), even though the scale of the critical axis (BAU) was 1000 times smaller than the scale used in (A,B); i.e., the scale facilitated the detection of any pattern, if present.

Figure 5

Three-dimensional analysis of age data. The age of SIRS+, infected individuals did not explain the mortality rates described in Figure 4. Both study I (A) and II (B) included >65-year-old individuals in the left (low-mortality) subset and <50-year-old individuals in the right (high-mortality) subset. While the median age [large, blue circles (C,D)] was 7–15 years higher in the right subset, any age-based cutoff would result in a large number of errors because the age intervals of survivors and non-survivors overlapped (C,D).

Figure 6

Assessment of redundancy (structure II). When all 105 SIRS+, infection+ individuals were assessed, two (“left” and “right”) perpendicular subsets were observed, which differed in mortality rates: it was 26.1% in the left subset and 54.2% in the right subset (P < 0.004, Chi-square test, Table 1). For clarity, the X and Y axes are scaled down, and three data points are not plotted (A). To prevent errors and improve the chances of extracting more information from the same data, an additional data structure was investigated, which also showed two data subsets orthogonal to one another (B). In contrast, SIRS-negative, non-infected individuals did not present the high-mortality pattern, even though the scale of the Y axis (BAU) was 1000 times smaller than the scale used when SIRS+, infection+ individuals were tested (C).

Figure 7

Study-specific, subset-specific assessment of mortality. When the second data structure was analyzed in each study, three spatial patterns were observed, both in study I (A) and II (B), which showed differences in mortality. When an additional group of seven septic (SIRS+, infection+) individuals was tested over time, three patients were distinguished at day 2 or 3, who displayed (i) a “right” pattern [observable at day 2, two patients (red oval)] or (ii) a “left” pattern [detected at day 3, one patient (purple square) (C)]. When time was not considered, spatial patterns identified three subsets, classified as “left,” “vertical,” and “right” (D).

Figure 8

Study-specific, sub-specific assessment of mortality (structure III). At least three profiles were distinguished when a third data structure was explored (A,B). Mortality differed up to five times across subsets. Longitudinal data of septic patients displayed similar patterns, distinguishing at least three individuals [red oval, purple square (C)].

Figure 9

Validation of subsets detected by structure II. The immune profiles of subsets detected in Figures 7A,B were investigated. Both study I (A) and study I (B) showed non-randomly distributed leukocyte profiles, both within- and between-subsets. For instance, in both populations, within-subset differences were observed in the “right” subset, where the lymphocyte and neutrophil percentages did not overlap between survivors and non-survivors [blue horizontal lines (A,B)]. Between-subset differences were also observed, e.g., “left” subset survivors displayed higher L%, higher M%, and lower N% than survivors classified within the remaining subsets [red horizontal lines (A,B)]. Horizontal lines show some data subsets that did not overlap.

Figure 10

Validation of subsets detected by structure III. The immune profiles of subsets detected in Figures 8A,B were investigated. Discrimination was repeatable, even when a different data structure was utilized. While patterns differed slightly [study I (n = 36) showed four subsets, and one pattern (“vertical”) was only composed of one outcome (non-survivors) (A)], study II (n = 69) detected three subsets, with two outcomes per subset (B). The “left” and “right” subsets of both studies reproduced the information observed in Figure 8: while the “left” subset did not show within-subset differences, the lymphocyte and neutrophil percentage intervals of the “right” subset did not overlap. The “vertical” subset of study II seemed to correspond to both the “vertical” and “lower left” subsets of study I. Horizontal lines show some data subsets that did not overlap.

Figure 11

Assessment of immune functions. Based on the subsets detected in Figures 7A,B, interactions involving two or three cell types were investigated. Both study I (A) and study II (B) conveyed similar information. Both survivors and non-survivors of the “left” subset displayed the lowest neutrophil/lymphocyte (N/L) and highest mononuclear cell/neutrophil (MC/N) values. The “right” subset showed the lowest monocyte/neutrophil (M/N) and monocyte/lympohocyte (M/L) values, as well as within-subset differences were not observed in the remaining subsets: “right” non-survivors revealed lower N/L, M/N, and M/L and higher MC/N values than “right” survivors. The “vertical” non-survivors showed the highest N/L and the lowest MC/N values of all non-survivors. Together with the information shown in Figures 9A,B, these patterns support three hypotheses (i) the “right” subset experienced a monocyte-mediated immunosuppression; (ii) the “vertical” subset expressed excessive inflammation, together with low lymphocyte percentages; and (iii) mortality was not due, in the “left” subset, to any of the three disorders observed in the other subsets.

Figure 12

Assessment of disease dynamics. The combinatorial approach also measured disease dynamics, expressed as temporal interactions that included antibiotic–microbial–immunological relationships. Temporal information (days after admission) was not informative when either one or two levels of interactions were measured [N/MC, M/L, and (M/L)/(N/MC) (A–C)]. One-level interactions also failed to discriminate when the spatial patterns shown by Figure 7D were considered (D,E). In contrast, non-overlapping data distributions were observed when both spatial profiles and two-level interactions were assessed, confirming the expectation that discrimination increases when two or more levels of complexity are investigated (F). Horizontal lines denote non-overlapping data distributions (F).