Using Digital Data to Predict CHD: Complete Statistical Analysis

1. Trial at a Glance

This page covers the registry results for Using Digital Data to Predict CHD (NCT04574882), a registered phase not stated study. The primary outcome is Latent Dirichlet Allocation (LDA) Topics - Topics / Themes Discussed Between Patients With and Without Heart Disease.

2. Clinical Question

The study, titled “Using Digital Data to Predict CHD,” seeks to identify digital features associated with coronary heart disease and related risk factors. Its broader aim is to develop models combining digital data with conventional prediction approaches for CHD risk and health care utilization.

The primary outcome description focuses on topics and features derived from language, rather than a clinical event comparison. It says that features were assessed for association with case status and describes analyses of cardiovascular risk prediction as well. The registry record does not give numerical results showing the strength or precision of those associations or prediction performance.

3. Trial Design

This completed study is registered as NCT04574882 and is sponsored by the University of Pennsylvania. The registry lists enrollment of 781. The study started on 2020-09-25, has a primary completion date of 2025-05-30, and has status COMPLETED.

The record describes a case and control structure rather than random assignment to treatments. The case description refers to patients ages 30–74 with and without CHD, with the condition identified within the last 5 years. The control description refers to patients aged 30–74 with a non-cardiovascular-related history. The record does not state how allocation to these groups occurred, and its notes caution that the arms are not a like-for-like control comparison.

A survey is listed as the intervention, with type OTHER. The study description centers on digital data, including social media, online search, and mobile media, and specifically describes use of available Facebook wall posts. It says that posts from account history through data collection were included regardless of whether they occurred before or after a CHD diagnosis. Thus, the registry describes an observational analysis of digital features, not a trial assigning a medical treatment.

The registry lists three outcome measures but zero analysis records. The primary completion timeframe is described as through study completion, an average of 3 years. The record does not report a detailed recruitment process, the timing of data collection for each participant, or a protocol-level account of the statistical analysis plan. It also does not report a sample-size justification. These gaps limit assessment of how the registered methods relate to all analyses ultimately conducted.

4. Participant Flow and Accounting

The registry reports how participants moved through each stage of the study. All counts below are exactly as posted; missing counts are not inferred from other stages.

Overall Study

Analysis population for the primary outcome, as defined in the registry: To evaluate predictive performance for ASCVD risk, we used regression and classification analyses. Pearson correlations assessed how language features predicted continuous risk scores. For classification, we measured AUC for binary risk (≥10% vs \<10%) and standard risk categories. Three logistic regression models were tested: language alone, demographics alone, and combined. Arms/groups were combined because the goal was to predict ASCVD risk continuously and categorically across participants.

5. Baseline Characteristics

The table reproduces the baseline characteristics posted in the registry, which describe the participants analysed at baseline and help show whether the randomized groups were comparable.

6. Primary Endpoint

The registry lists the primary outcome as Latent Dirichlet Allocation (LDA) Topics - Topics / Themes Discussed Between Patients With and Without Heart Disease, measured over Through study completion, an average of 3 years.

The primary outcome is titled “Latent Dirichlet Allocation (LDA) Topics - Topics / Themes Discussed Between Patients With and Without Heart Disease.” Its description characterizes the outcome as topics and features derived by clustering language data. In the stated approach, LDA identifies themes from co-occurring words, and each language feature is encoded as a normalized frequency count per user. The resulting measures can be examined for association with CHD case status.

The registry also describes unigrams, LIWC categories, and LDA topics as feature types. It says separate logistic regression models and Pearson correlations were used to assess predictive value for case status. A separate description addresses ASCVD risk, including risk scores and binary risk categories. These are related analytic aims, but the registry does not provide numerical findings that clarify how each contributes to the primary outcome.

A language feature can serve as a candidate marker because patterns in words or themes may differ across groups. Correlation describes the direction and strength of an association between variables; logistic regression relates features to a binary status. Neither, by itself, establishes that language causes disease or that a model will predict accurately in new participants.

The outcome metadata are incomplete: the unit is truncated after “AUC (Area Under t,” while dispersion is labelled “100% Confidence Interval.” No estimate, interval bound, or feature-level result is reported in the registry record. Consequently, the intended endpoint can be described, but its observed magnitude and precision cannot be evaluated from the registry information.

7. Why No Re-analysis Is Shown

8. Statistical Interpretation

9. Statistical Methods Explained

The registry describes language processing and statistical association methods intended to characterize digital features in relation to heart disease status and cardiovascular risk. It does not include analysis records or numerical estimates. The concepts below explain what the named methods generally do; they do not establish how well any method performed in this study.

What does LDA do?

Latent Dirichlet Allocation is a topic-modeling approach that represents documents as mixtures of topics and topics as patterns of words. In general, it can be used to identify recurring themes in a collection of text without first assigning each post to a pre-defined clinical category. The registry says that LDA was used to generate 200 clusters of co-occurring words, described as “topics.” A topic is a statistical summary of word co-occurrence, not automatically a validated clinical construct. Interpretation depends on the words that characterize it and on how consistently it can be identified.

The description says that all available Facebook wall posts from the start of each account history through data collection were included, whether posted before or after CHD diagnosis. It also says that language features were encoded as normalized frequency counts per user. Normalization can make feature values more comparable across users with different amounts of text, but it does not eliminate all differences in posting behavior, account history, or text availability.

What do Pearson correlation and logistic regression assess?

Pearson correlation describes the direction and strength of a linear relationship between two measured variables. In this setting, the record says correlations were used to assess linguistic features in relation to case status and, in a separate risk-prediction description, continuous risk scores. Correlation is a measure of association. It does not show that a feature causes a health outcome, and it may not capture relationships that are not linear.

Logistic regression is generally used when the outcome is binary, such as presence versus absence of a condition. It models how the odds of the outcome relate to one or more predictors. The registry describes separate logistic regression models for each feature type, including unigrams, LIWC categories, and LDA topics. The record does not give coefficient estimates, model specifications beyond this description, or information about adjustment for other predictors. It is therefore not possible to assess the size or stability of any feature association.

What does AUC mean?

For a classification model, the area under the receiver operating characteristic curve (AUC) is a measure of discrimination: generally, how well model scores rank people with and without the outcome across possible thresholds. It does not by itself show whether predicted probabilities are well calibrated, whether a particular decision threshold is useful, or whether performance will transfer to a different population. The registry mentions AUC for binary risk classification, but the endpoint unit is truncated and no AUC result is reported.

For a binary risk definition, the record refers to a threshold of ≥10% versus <10% and to standard risk categories. The registry also describes Pearson correlations for continuous risk scores. These approaches address different questions: a correlation concerns covariation with a continuous value, while classification performance concerns separating groups defined by a category. Neither result is reported numerically in the registry record.

Why does validation matter?

Prediction models may appear more effective when assessed using the same data that informed their development. General statistical practice therefore distinguishes model development from evaluation in data not used to fit the model. The registry record does not report a validation strategy, data partitioning, external validation, or estimates of performance on new data. No assumption should be made about whether these steps were conducted.

Feature selection and topic generation can also affect apparent associations. If many language features are examined, some may appear associated by chance, particularly without an appropriate multiplicity strategy. The record describes multiple feature types and 200 LDA topics, but does not report a multiplicity adjustment or a prespecified procedure for selecting features. That uncertainty limits interpretation of any individual association even if estimates were available.

Over-dispersion is a general concern when count-like observations vary more than a simple statistical model expects. Normalized word frequencies are not raw counts, and the registry does not describe a count model or an over-dispersion assessment. It would therefore be inappropriate to infer that over-dispersion was present or handled in a particular way. More broadly, the stated methods identify analytic approaches, while the registry record does not report enough model detail or numerical output to assess their assumptions or performance.

10. Multiplicity, Missing Data and Analysis Populations

The registry describes several classes of language features: unigrams, LIWC categories, and LDA topics. It also describes Pearson correlations and logistic regression, with analyses concerning CHD status and ASCVD risk. When many features or related outcomes are examined, the chance of finding apparently notable associations by chance can increase. General statistical practice may address this through prespecified hypotheses, adjustment for multiple comparisons, or emphasis on validation. The registry record does not report a multiplicity strategy, a hierarchy of hypotheses, or which features were selected for reporting.

The record does not report a missing-data strategy. It says that all available Facebook wall posts were included, but does not state how unavailable accounts, absent posts, incomplete surveys, or missing clinical information were handled. Availability of digital text may differ across participants, and normalized frequency counts do not by themselves explain how missing or sparse information was treated.

Analysis populations are not reported in detail. The registry lists enrollment of 781 and describes case and control groups, but does not give analysis denominators for the primary outcome or risk analyses. It does not say whether all enrolled participants contributed to each analysis, how eligibility for specific analyses was determined, or whether exclusions occurred.

The registry lists zero analysis records and does not report a sample-size justification, interim analyses, or a detailed statistical analysis plan. These items are not reported in the registry record; their absence from the record does not establish what may or may not have occurred outside the available entry.

11. What the Result Does, and Does Not, Mean

The registry describes a study intended to use digital information and language features to examine cardiovascular health, CHD status, and health care utilization. Its primary outcome description focuses on themes identified from text and their relationship to heart disease status. The methods mention Pearson correlation, logistic regression, and AUC-based classification assessment.

Those descriptions do not establish that a particular language feature was associated with CHD, that prediction was accurate, or that digital information improved on conventional models. The registry provides no numerical primary-outcome estimate, no AUC result, and no interval bounds. The analysis records count is zero. As a result, the magnitude and uncertainty of the findings cannot be assessed from the registry record.

Even a reported association would not by itself demonstrate causation. The study is not described as randomized, the allocation process is not stated, and the case and control groups are not presented as a like-for-like treatment comparison. The included Facebook posts may be from before or after CHD diagnosis, so an association would not necessarily show that language patterns preceded the condition.

Likewise, a model’s performance within the data used for its development would not establish how it performs in other people or settings. The registry record does not report validation details, calibration, or numerical predictive performance. It also does not provide enough information to assess the effect of missing digital data or feature selection.

The appropriate interpretation is therefore limited: the registry records a completed study with stated aims and methods, but does not report the results needed to judge associations, prediction, or precision. It should not be taken as evidence for or against a specific clinical prediction claim.

12. Why This Trial Matters Statistically

This study raises an important statistical question for digital health research: whether information collected outside conventional clinical measurements can be represented as reproducible features and evaluated as a predictor of cardiovascular outcomes. Its focus on text themes and normalized feature frequencies illustrates how unstructured digital material can be converted into variables that statistical models can analyze.

The analysis also highlights the distinction between association and prediction. A language feature may correlate with case status without improving prediction, while a model’s discrimination does not necessarily imply accurate probability estimates or useful decisions. Separate evidence is needed to assess these properties, and performance in data used to develop a model may not generalize. The registry record does not report numerical results or validation details, so these questions remain unanswered by the entry.

Study design and timing shape what such analyses can mean. The case and control groups are observational, their allocation is not stated, and language data may include posts from before and after diagnosis. These features constrain causal and temporal interpretations. A digital signal could reflect many aspects of participants’ lives or the experience of illness; the registry record does not establish a causal pathway.

Finally, examining many language features brings attention to multiplicity, feature selection, and reproducibility. Topic models can yield useful summaries, but a topic’s statistical construction does not guarantee a stable or clinically interpretable measure. Reporting model details, uncertainty, missing-data handling, and validation is necessary for readers to evaluate whether an observed pattern is robust. The registry entry states the research aims and selected methods, but does not provide the outcome information required for that evaluation.

13. Limitations

• The posted summary statistics are the only data these results rest on; patient-level data, full protocols and statistical analysis plans are not available.
• Recalculated results are simple, unadjusted calculations from posted summaries. They do not replace the sponsor's prespecified analysis.

The main limitation for interpretation is the absence of numerical outcome results. The registry lists three outcome measures but zero analysis records, and it does not report the primary endpoint estimate or confidence interval bounds. The endpoint metadata are also incomplete: the unit text is truncated, and the stated “100% Confidence Interval” label cannot be interpreted without corresponding numerical limits.

The primary outcome description combines topic generation, feature association, and predictive performance. It does not make clear in the registry record how these elements were translated into a single reported endpoint. The broader description mentions both CHD status and ASCVD risk, but no numerical output is reported for either. The stated methods therefore cannot be evaluated against observed findings.

The registry gives limited information about design and analysis populations. It describes case and control groups, but allocation is not stated and the record cautions that the arms are not a like-for-like control comparison. It does not provide analysis denominators, participant flow, exclusions, missing-data handling, or a detailed account of which participants contributed to each analysis.

The use of posts from before and after diagnosis complicates temporal interpretation. The registry does not report a separate analysis by timing relative to diagnosis. It also does not report how digital-data availability varied across participants, whether the text features were externally validated, or whether the topic structure was assessed for stability. These omissions do not prove that such analyses were not performed; they are not reported in the registry record.

Finally, the entry does not report a multiplicity strategy or sample-size justification. Together with the lack of estimates and validation information, this prevents assessment of statistical precision, robustness, and generalizability.

14. Related Tutorials

15. Related Statistical Calculators

16. Sources

• ClinicalTrials.gov: Using Digital Data to Predict CHD — NCT04574882.
• Registry record last updated 2025-11-04; status last verified 2025-10.
• Results first posted 2025-11-04.
• Data retrieved from ClinicalTrials.gov on 2026-11-09.
• Group sizes, means, event counts and sponsor-reported p-values are reproduced as posted. Effect estimates, confidence intervals and p-values not attributed to the sponsor are independently calculated from the posted summary statistics.

17. Record Summary

NCT04574882, “Using Digital Data to Predict CHD,” is sponsored by the University of Pennsylvania and is listed as COMPLETED. The study started on 2020-09-25, has a primary completion date of 2025-05-30, and lists enrollment of 781. The registry describes a survey and a case/control structure involving patients aged 30–74.

The study’s aims include identifying digital features associated with CHD and related risk factors and developing prediction models for CHD risk and health care utilization. Its primary outcome description concerns LDA-derived language topics and features. The methods description mentions normalized language-feature frequencies, Pearson correlation, logistic regression, and AUC assessment.

The registry lists three outcome measures and zero analysis records. Numerical primary-outcome results, uncertainty estimates, detailed analysis populations, and validation results are not reported in the registry record. The entry therefore documents study aims and methods, but does not permit a quantitative assessment of the reported prediction or association performance.