Multivariate Functional Principal Component Analysis for Mixed-Type mHealth Data: An Application to Mood Disorders

This paper proposes a semiparametric multivariate functional principal component analysis method for mixed-type mHealth data, which effectively identifies interpretable time-of-day patterns across diverse health domains to stratify mood disorder subtypes.

The Gist

Imagine you are trying to understand the daily life of a person with a mood disorder. You have two types of information streaming in:

1. The "Feelings" Stream: Every few hours, the person checks a phone app and rates their sadness, anxiety, and energy on a scale of 1 to 7. These are ordinal numbers (rankings).
2. The "Movement" Stream: A watch on their wrist records their physical activity every minute. This is continuous data (a smooth flow of numbers).

The problem is that these two streams are messy. The feelings are checked at random times, the numbers are different types (ranks vs. raw counts), and they are all mixed together. Traditional statistical tools are like a chef who only knows how to cook with fresh vegetables; they get confused when you throw in a raw egg, a block of cheese, and a jar of pickles all at once. They can't make sense of the whole meal.

This paper introduces a new "kitchen" called M2FPCA (Multivariate Functional Principal Component Analysis for Mixed-Type data). Here is how it works, using simple analogies:

1. The "Secret Language" Analogy (The Core Idea)

The authors realized that even though the data looks different (ranks, yes/no, counts), it all comes from a single, hidden "secret language" inside the person's brain and body.

• The Metaphor: Imagine a complex orchestra playing music. To the audience, you hear violins (continuous), drums (binary on/off), and a choir singing in different volumes (ordinal). It sounds chaotic.
• The Solution: The authors propose that there is a single conductor (a hidden Gaussian process) directing the whole orchestra. If we can figure out what the conductor is doing, we can understand the whole symphony, even if the instruments are different.
• The Magic Trick: They use a mathematical "translator" (a Gaussian Copula) to convert all the messy, different types of data into this single, smooth "conductor's score." Now, the computer can analyze the whole day's pattern as one unified story.

2. The "Shadow Puppet" Analogy (Handling Missing Data)

In real life, people forget to check their phones, or their watches run out of battery. The data has holes.

• The Metaphor: Imagine looking at a shadow puppet show on a wall. Sometimes the puppet's hand is missing because the light flickered.
• The Solution: Because the authors know the "conductor's score" (the underlying pattern), they can use the parts of the shadow that are visible to guess what the missing parts looked like. They don't just ignore the holes; they fill them in based on the rhythm of the rest of the show.

3. The "Two Ways to Listen" (The Two Methods)

The paper offers two ways to analyze this data, depending on how much computing power you have:

• Method A (M2FPCA): The "Super-Computer" Approach.
  This method listens to every single instrument and how they talk to every other instrument simultaneously. It's incredibly detailed and accurate, like a high-end sound engineer analyzing every frequency. However, it's very slow and heavy if you have a huge orchestra (lots of data types).
• Method B (ps-M2FPCA): The "Efficient" Approach.
  This method assumes that all instruments share a few main "rhythms" (principal components). It's like saying, "The whole band is mostly following the same beat, but with slight variations." It's much faster and lighter, perfect for large studies, and the authors found it works almost as well as the heavy method for their specific data.

4. The Real-World Discovery (The "Aha!" Moment)

The authors tested this on 307 people with mood disorders. They didn't just look at "how sad" someone was; they looked at the shape of their day.

They found three main "Daily Patterns" (Digital Biomarkers) that act like fingerprints for different disorders:

1. The "Burden" Pattern: A general level of how tired, anxious, and inactive a person feels all day.
2. The "Morning-to-Night" Pattern: A natural rhythm where energy drops and anxiety rises as the day goes on (like a battery draining).
3. The "Midday Peak" Pattern: A bump in energy or mood right around lunchtime.

The Big Reveal:

• Bipolar Disorder patients showed a very specific "Morning-to-Night" pattern in their physical activity that was different from others.
• Major Depression patients had a "flat" energy curve—they didn't have that natural morning-to-night drop; they just felt low all day.
• Anxiety was the common thread that showed up in almost everyone, but the way it moved through the day helped tell the disorders apart.

Why This Matters

Before this, doctors might have looked at a patient's "average sadness" and "average activity" and missed the story. This method is like watching a movie of the patient's day instead of just looking at a snapshot.

It allows researchers to say: "It's not just that you are sad; it's that your sadness and energy get out of sync with your body's natural clock in a specific way that tells us exactly which type of mood disorder you have." This is a giant step toward precision psychiatry, where treatment is tailored to the specific "rhythm" of a person's brain.

Technical Summary

Here is a detailed technical summary of the paper "Multivariate Functional Principal Component Analysis for Mixed-Type mHealth Data: An Application to Mood Disorders."

1. Problem Statement

Modern mobile health (mHealth) studies generate high-frequency, multimodal data combining self-reported ecological momentary assessments (EMA) with passively collected sensor data (e.g., actigraphy). These datasets present significant statistical challenges:

• Mixed Measurement Scales: Data includes continuous (physical activity), truncated (pain), ordinal (mood, energy), and binary (events) variables.
• Irregular Sampling: Observations occur at irregular times within the day and vary across subjects.
• Complex Dependence: There is intricate temporal dependence within variables and cross-dependence between different behavioral domains (e.g., how mood correlates with activity over time).
• Limitations of Existing Methods: Standard Functional Principal Component Analysis (FPCA) assumes continuous Gaussian data. Existing generalized FPCA methods often focus on regression settings or univariate outcomes, failing to provide a coherent joint model for multivariate mixed-type functional data.

2. Methodology: M2FPCA Framework

The authors propose Multivariate Functional Principal Component Analysis for Mixed-Type Data (M2FPCA), built upon a Multivariate Generalized Latent Nonparanormal Process (MGLNPP) model.

A. Data-Generating Mechanism (MGLNPP)

The model assumes observed mixed-type data $X(\cdot)$ arises from a latent multivariate Gaussian process $V(\cdot)$ through a three-step hierarchy:

1. Latent Gaussian Process: A mean-zero $J$-variate Gaussian process $V(\cdot)$ with a correlation kernel $C_{jk}(s,t)$ induces dependence.
2. Monotone Transformations: Component-wise monotone functions transform $V$ into a latent continuous process $Z$.
3. Type-Specific Observation Maps: The latent process $Z$ is mapped to observed data $X$ via specific mechanisms:
   • Continuous: Identity mapping.
   • Truncated: Thresholding (e.g., $X = Z \cdot I\{Z > \Delta\}$).
   • Ordinal: Binning via cutoffs (e.g., $X = k$ if $\Delta_k \le Z < \Delta_{k+1}$).
   • Binary: Thresholding (e.g., $X = I\{Z > \Delta\}$).

B. Estimation Strategy

Since observed data are non-Gaussian, direct covariance estimation is invalid. The authors use Kendall's $\tau$ bridging, which is invariant to monotone transformations, to estimate the latent correlation structure.

• Step 1: Latent Correlation Estimation. Sample Kendall's $\tau$ statistics are computed for pairs of observations. These are mapped to the latent correlation surface $C_{jk}(s,t)$ using type-specific bridge functions derived from the copula model.
• Step 2: Covariance Surface Modeling. The correlation surfaces are modeled using tensor-product splines.
• Step 3: Latent Prediction. Best Linear Unbiased Predictors (BLUP) are used to estimate subject-specific latent Gaussian trajectories $\hat{V}_i(t)$ given the observed data and estimated covariance.

C. Two Estimation Procedures

1. M2FPCA (Fully Multivariate):
   • Jointly estimates all marginal and cross-correlation surfaces $\{C_{jk}\}$.
   • Performs multivariate FPCA on the predicted latent trajectories.
   • Pros: Captures full dependence structure. Cons: Computationally expensive ($O(J^3|S|^3)$).
2. ps-M2FPCA (Partially Separable):
   • Assumes a partial separability structure where temporal eigenfunctions are shared across domains, but score covariances vary.
   • Stage I: Estimates univariate latent trajectories using M1FPCA.
   • Stage II: Applies multivariate FPCA to the pooled latent predictions.
   • Pros: Computationally scalable ($O(J|S|^3)$) and suitable for moderate-to-large numbers of domains.

3. Key Contributions

1. Unified Probabilistic Framework: Introduction of the MGLNPP, which unifies continuous, truncated, ordinal, and binary functional data under a single latent Gaussian copula model.
2. Semiparametric Estimation: Development of a robust estimation procedure using Kendall's $\tau$ bridging to handle mixed data types without assuming specific marginal distributions.
3. Scalable Algorithms: Proposal of two estimators (M2FPCA and ps-M2FPCA), with the latter offering a computationally efficient alternative for high-dimensional functional data.
4. Digital Biomarkers: The method yields interpretable latent functional principal component (FPC) scores that serve as participant-specific digital biomarkers summarizing coordinated dynamics across multiple behavioral domains.

4. Results

A. Simulation Studies

• Setup: Simulated 4-dimensional mixed-type data (binary, ordinal, truncated, continuous) under stationary and non-stationary covariance structures.
• Performance:
  • M2FPCA significantly outperformed standard MFPCA (designed for continuous data) and the partially separable ps-M2FPCA in non-separable scenarios, reducing Integrated Squared Error (ISE) by ~6x to 12x.
  • ps-M2FPCA performed comparably to M2FPCA in scenarios with partial separability and was vastly superior to standard MFPCA (~20x lower ISE).
  • Both proposed methods showed consistent improvement as sample size increased.

B. Application to NIMH Family Study

• Data: 307 participants (Controls, MDD, Bipolar I, Bipolar II) with 14 days of actigraphy and 4 daily EMA prompts (Sad Mood, Anxiousness, Energy, Activity).
• Findings:
  • Latent Structure: The top 3 FPCs explained 82% of variability.
    • FPC1 (64%): Overall daily burden of symptoms and activity.
    • FPC2 (12%): Monotone morning-to-evening contrast (homeostatic sleep pressure).
    • FPC3 (6%): Inverted-U pattern peaking at midday (circadian influence).
  • Clinical Stratification: Using LASSO regression on latent FPC scores to classify disorders:
    • Anxiousness: Mean and variability of FPC1/FPC2 were broadly informative for all mood disorders.
    • Bipolar Disorders: Showed distinct patterns in physical activity (TLAC) homeostatic gradients. Bipolar I had negative associations with activity mean, while Bipolar II showed positive associations.
    • MDD: Uniquely linked to reduced variability in energy (FPC1) and a weaker within-day energy gradient (FPC2).
    • Sad Mood: Did not emerge as a dominant discriminator, suggesting that dynamic interactions across the full mood circumplex are more critical than sadness severity alone.

5. Significance and Conclusion

• Methodological Advance: This work bridges the gap between functional data analysis and mixed-type data, providing a rigorous framework for the complex, irregular, and heterogeneous data typical of modern mHealth.
• Clinical Utility: The identified latent components (e.g., homeostatic gradients, circadian peaks) offer mechanistic insights into mood disorders. The resulting digital biomarkers can effectively stratify mood disorder subtypes, moving beyond static symptom checklists to dynamic, time-resolved phenotyping.
• Future Directions: The authors suggest extending the framework to multilevel models (M3FPCA) to account for serial dependence across days and validating these biomarkers across diverse cohorts and devices.

In summary, the paper presents a robust, flexible, and computationally feasible statistical framework for analyzing complex mHealth data, demonstrating its ability to uncover clinically meaningful patterns that traditional methods would miss.