Limits...
Identification of complex metabolic states in critically injured patients using bioinformatic cluster analysis.

Cohen MJ, Grossman AD, Morabito D, Knudson MM, Butte AJ, Manley GT - Crit Care (2010)

Bottom Line: While patients transitioned between states, they spent significant amounts of time in each.Further analysis of correlations between pairs of variables within each cluster reveals significant differences in physiology between clusters.These results demonstrate that hierarchical clustering techniques can be useful for visualizing complex multivariate data and may provide new insights for the care of critically injured patients.

View Article: PubMed Central - HTML - PubMed

Affiliation: Department of Surgery, University of California, 505 Parnassus Avenue, San Francisco, CA 94143, USA. mcohen@sfghsurg.ucsf.edu

ABSTRACT

Introduction: Advances in technology have made extensive monitoring of patient physiology the standard of care in intensive care units (ICUs). While many systems exist to compile these data, there has been no systematic multivariate analysis and categorization across patient physiological data. The sheer volume and complexity of these data make pattern recognition or identification of patient state difficult. Hierarchical cluster analysis allows visualization of high dimensional data and enables pattern recognition and identification of physiologic patient states. We hypothesized that processing of multivariate data using hierarchical clustering techniques would allow identification of otherwise hidden patient physiologic patterns that would be predictive of outcome.

Methods: Multivariate physiologic and ventilator data were collected continuously using a multimodal bioinformatics system in the surgical ICU at San Francisco General Hospital. These data were incorporated with non-continuous data and stored on a server in the ICU. A hierarchical clustering algorithm grouped each minute of data into 1 of 10 clusters. Clusters were correlated with outcome measures including incidence of infection, multiple organ failure (MOF), and mortality.

Results: We identified 10 clusters, which we defined as distinct patient states. While patients transitioned between states, they spent significant amounts of time in each. Clusters were enriched for our outcome measures: 2 of the 10 states were enriched for infection, 6 of 10 were enriched for MOF, and 3 of 10 were enriched for death. Further analysis of correlations between pairs of variables within each cluster reveals significant differences in physiology between clusters.

Conclusions: Here we show for the first time the feasibility of clustering physiological measurements to identify clinically relevant patient states after trauma. These results demonstrate that hierarchical clustering techniques can be useful for visualizing complex multivariate data and may provide new insights for the care of critically injured patients.

Show MeSH

Related in: MedlinePlus

Probability of infection in each cluster. The baseline infection rate (dashed line) is 0.735. There were two enriched for infection. Clusters 3 and 7 had too few data points for the proportions to be meaningful.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC2875524&req=5

Figure 3: Probability of infection in each cluster. The baseline infection rate (dashed line) is 0.735. There were two enriched for infection. Clusters 3 and 7 had too few data points for the proportions to be meaningful.

Mentions: We next examined the states produced from the clustering to determine if any of the clusters represented physiology that would be obvious to an astute clinician. We enumerated the physiological state of each cluster by calculating the means and standard deviations of each of the variables of the clusters (Table 2). Evaluation of the clinical data in these states by four experienced clinicians (intensivists and surgeons) resulted in an inability to clinically define any of the states as sick or well, resuscitated or unresuscitated, and so on, highlighting the difficulty of deriving any traditional clinical prediction or meaning from these patterns. Specifically, none of the clinicians were able to determine whether cluster x represented under resuscitation or cluster y was that of a well resuscitated patient. Because the clustering method failed to separate the patient data into groups by obvious traditional physiological definitions these results confirm our hypothesis that clustering would find meaningful patterns of data that were otherwise impossible to physiologically discern or classify using traditional clinical definitions. We next sought to test the predictive ability of our clustering method by calculating the distribution of patients with particular outcomes across the clusters. This was done for three outcomes: mortality, multiple organ failure (MOF), and infection. Briefly, the percentage of data points in each cluster that were from patients with a given outcome was calculated for each of the three outcomes. A baseline for comparison was calculated by dividing the total number of measurements across the whole data set from patients with a particular outcome by the total number of data points. Figure 2 shows that the baseline number of data points in the entire dataset from patients that died was 10.8%. Three clusters (2, 4, and 5) had higher representation of physiology correlated with death than baseline. Others had an underrepresentation of patients who died (clusters 1, 6, and 10). This was repeated for MOF and infection. Even with increasing baseline values (MOF = 0.47, infection = 0.73) there were six clusters that were enriched for MOF and two enriched for infection (Figures 3 and 4).


Identification of complex metabolic states in critically injured patients using bioinformatic cluster analysis.

Cohen MJ, Grossman AD, Morabito D, Knudson MM, Butte AJ, Manley GT - Crit Care (2010)

Probability of infection in each cluster. The baseline infection rate (dashed line) is 0.735. There were two enriched for infection. Clusters 3 and 7 had too few data points for the proportions to be meaningful.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC2875524&req=5

Figure 3: Probability of infection in each cluster. The baseline infection rate (dashed line) is 0.735. There were two enriched for infection. Clusters 3 and 7 had too few data points for the proportions to be meaningful.
Mentions: We next examined the states produced from the clustering to determine if any of the clusters represented physiology that would be obvious to an astute clinician. We enumerated the physiological state of each cluster by calculating the means and standard deviations of each of the variables of the clusters (Table 2). Evaluation of the clinical data in these states by four experienced clinicians (intensivists and surgeons) resulted in an inability to clinically define any of the states as sick or well, resuscitated or unresuscitated, and so on, highlighting the difficulty of deriving any traditional clinical prediction or meaning from these patterns. Specifically, none of the clinicians were able to determine whether cluster x represented under resuscitation or cluster y was that of a well resuscitated patient. Because the clustering method failed to separate the patient data into groups by obvious traditional physiological definitions these results confirm our hypothesis that clustering would find meaningful patterns of data that were otherwise impossible to physiologically discern or classify using traditional clinical definitions. We next sought to test the predictive ability of our clustering method by calculating the distribution of patients with particular outcomes across the clusters. This was done for three outcomes: mortality, multiple organ failure (MOF), and infection. Briefly, the percentage of data points in each cluster that were from patients with a given outcome was calculated for each of the three outcomes. A baseline for comparison was calculated by dividing the total number of measurements across the whole data set from patients with a particular outcome by the total number of data points. Figure 2 shows that the baseline number of data points in the entire dataset from patients that died was 10.8%. Three clusters (2, 4, and 5) had higher representation of physiology correlated with death than baseline. Others had an underrepresentation of patients who died (clusters 1, 6, and 10). This was repeated for MOF and infection. Even with increasing baseline values (MOF = 0.47, infection = 0.73) there were six clusters that were enriched for MOF and two enriched for infection (Figures 3 and 4).

Bottom Line: While patients transitioned between states, they spent significant amounts of time in each.Further analysis of correlations between pairs of variables within each cluster reveals significant differences in physiology between clusters.These results demonstrate that hierarchical clustering techniques can be useful for visualizing complex multivariate data and may provide new insights for the care of critically injured patients.

View Article: PubMed Central - HTML - PubMed

Affiliation: Department of Surgery, University of California, 505 Parnassus Avenue, San Francisco, CA 94143, USA. mcohen@sfghsurg.ucsf.edu

ABSTRACT

Introduction: Advances in technology have made extensive monitoring of patient physiology the standard of care in intensive care units (ICUs). While many systems exist to compile these data, there has been no systematic multivariate analysis and categorization across patient physiological data. The sheer volume and complexity of these data make pattern recognition or identification of patient state difficult. Hierarchical cluster analysis allows visualization of high dimensional data and enables pattern recognition and identification of physiologic patient states. We hypothesized that processing of multivariate data using hierarchical clustering techniques would allow identification of otherwise hidden patient physiologic patterns that would be predictive of outcome.

Methods: Multivariate physiologic and ventilator data were collected continuously using a multimodal bioinformatics system in the surgical ICU at San Francisco General Hospital. These data were incorporated with non-continuous data and stored on a server in the ICU. A hierarchical clustering algorithm grouped each minute of data into 1 of 10 clusters. Clusters were correlated with outcome measures including incidence of infection, multiple organ failure (MOF), and mortality.

Results: We identified 10 clusters, which we defined as distinct patient states. While patients transitioned between states, they spent significant amounts of time in each. Clusters were enriched for our outcome measures: 2 of the 10 states were enriched for infection, 6 of 10 were enriched for MOF, and 3 of 10 were enriched for death. Further analysis of correlations between pairs of variables within each cluster reveals significant differences in physiology between clusters.

Conclusions: Here we show for the first time the feasibility of clustering physiological measurements to identify clinically relevant patient states after trauma. These results demonstrate that hierarchical clustering techniques can be useful for visualizing complex multivariate data and may provide new insights for the care of critically injured patients.

Show MeSH
Related in: MedlinePlus