In this case study, I describe how Dr.DataScience used anonymized clinical-research data to carry out a complex data analysis and help the client solve a clinical question. In accordance with our confidentiality agreement, no specific disease names, measured items, figures, individual patient information, or institution names are disclosed; however, the analytical methods used, how they were applied, and the type of findings obtained are the same as in the actual analysis.

Background and objective

The aim of this case was to identify factors—such as multiple physiological indicators, patient background information, and treatment-related data—that might influence whether a particular clinical outcome occurs, and to clarify their complex relationships. In particular, the client’s question was “I want to understand what conditions influence the clinical outcome,” and the task was to provide objective evidence through statistical analysis so as to obtain a deeper understanding and clear insights.

Data and variables

The analysis used multiple laboratory values and biomarkers contained in the patients’ clinical data, treatment-related information, and the response variable, namely the presence or absence of a particular clinical outcome.

Analytical methods

1. Test of normality (Shapiro–Wilk test)
   • Performed to confirm whether the distribution of each variable followed a normal distribution. Because normality was not observed for any of the variables analyzed, I decided to apply nonparametric methods.
2. Correlation analysis (Spearman’s rank correlation coefficient)
   • To grasp the nonparametric correlations among the variables, I conducted a correlation analysis using Spearman’s rank correlation coefficient. This confirmed statistically significant correlations among many of the variables.
3. Decision-tree analysis (CART) and detailed group analysis with binomial tests
   • The client asked me to identify not a single indicator, but finer classification criteria that would better support judgment in the clinical setting. For this reason, rather than ordinary logistic regression, I performed decision-tree analysis (CART), which can clearly separate, at split points (thresholds), the patterns of factors that influence the response variable, “presence or absence of the clinical outcome.”

     Through this analysis, I hierarchically identified, from among multiple explanatory variables, those with a strong influence on the response variable, and divided the data into groups by conditional branching.

     
     (Figure: Example of a decision tree [CART])
   • (a) Model considering the principal factors

     First, I ran the analysis using all relevant variables as explanatory variables, which showed that a certain principal indicator was the most influential factor. It was confirmed that, around a specific threshold of this indicator, the tendency for the clinical outcome to be present or absent diverged strongly.

   • (b) Model excluding some of the principal factors

     Because the principal indicator above was strongly correlated with other variables, I built a model excluding this indicator and explored different patterns of factors. This revealed that other laboratory values and patient background factors (e.g., BUN, dosage of a particular drug, Cr, etc.) influence the presence or absence of the clinical outcome at specific thresholds (e.g., BUN of 40, dosage of 1000, etc.).

   • (c) Model excluding still more factors

     By further building a model that also excluded a specific laboratory value (e.g., BUN), I analyzed the patterns predicting the presence or absence of the clinical outcome through the relationships among the remaining indicators (e.g., eGFR, the ratio of a particular laboratory value, dosage, etc.). Here too, specific thresholds of each indicator (e.g., eGFR of 35, a laboratory-value ratio of 25 or 65, etc.) served as the principal branch points influencing the presence or absence of the clinical outcome.

   • (d) Binomial test (with FDR correction)
     • For each group (node) obtained from the decision-tree analysis, I performed binomial tests to evaluate whether there was a statistically significant difference in the probability of the particular clinical outcome. To address the problem of multiple comparisons, I applied FDR (False Discovery Rate) multiple-comparison correction. Because this was an exploratory analysis, rather than a stricter correction such as the Holm method, I adopted FDR, which tolerates false positives to a certain level while being more likely to detect a greater number of associations.
     • This made it possible to objectively evaluate the “incidence of the clinical outcome” in each group classified by the decision tree, and to clearly show whether the difference was significant at the statistical significance level (e.g., α = 5%). For example, even under a particular condition (e.g., when a certain laboratory value was 36.5 or below and a certain ratio was greater than 25), the clinical outcome was present in 86.7% of cases, and a statistically significant difference was confirmed.

Overview of the main results and clinical considerations

This analysis showed that the presence or absence of the particular clinical outcome involves multiple factors in a complex interplay. In particular, I discovered patterns in which groups at high or low risk of the clinical outcome were clearly classified according to the thresholds of specific indicator values identified by the decision-tree analysis, and their combinations.

1. • Direct influence of the principal indicator

     The decision-tree analysis showed that the threshold of a specific indicator was most strongly associated with the presence or absence of the clinical outcome.

   • Involvement of compound factors

     In the analysis excluding the principal indicator, it became clear that the presence or absence of the clinical outcome is influenced by combinations of multiple different laboratory values, medication information, and patient background factors. For example, the client’s hypothesis that a higher ratio of a particular laboratory value is associated with greater involvement in the clinical outcome was supported by these results. I also confirmed that the handling of certain laboratory values is complex, and that there are patterns in which the probability of the clinical outcome rises in stages when specific conditions overlap.

These results provide important insights for assessing risk from multiple perspectives in clinical diagnosis and treatment-strategy decisions.

Dr.DataScience’s contribution

In this case, to address the client’s question—“I want to understand what conditions influence the clinical outcome”—and their request to identify finer classification criteria that go beyond a single indicator and better support clinical judgment, I designed and executed an analytical strategy centered on decision-tree analysis rather than ordinary regression.

By selecting and applying statistical methods suited to the characteristics of the data (confirmation of normality with the Shapiro–Wilk test, Spearman’s rank correlation, multi-stage decision-tree analysis, and binomial tests for each group classified by the decision tree), I provided objective evidence addressing the client’s question.

By extracting meaningful patterns from complex data, along with the specific differences in probability associated with them, and statistically testing the client’s hypotheses, I delivered a deeper clinical understanding and solid insights.