📄 medicine

Weak health phenotypes but distinct depressive-symptom subtypes among United States adults: an unsupervised analysis of NHANES 2007–2018

This unsupervised analysis of NHANES 2007–2018 data reveals that while broad health phenotypes among US adults are statistically distinct yet weakly separated and primarily driven by age, depressive symptoms exhibit a clearer, more complex subtype structure that suggests total scores may overlook clinically significant patterns.

Original authors: Muhammad Ikhlas Malik, Wong Qi En

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad Ikhlas Malik, Wong Qi En

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery about how people feel. In the world of mental health research, there is a very popular tool called the PHQ-9. Think of it like a nine-question quiz where people rate things like "feeling down," "trouble sleeping," or "feeling tired." Usually, scientists just add up all the points to get a single number. If the number is high, they say, "This person is depressed." It's like grading a test: a score of 90 means an A, and a score of 40 means an F.

But here is the problem with that simple math: two people can get the exact same score for totally different reasons. One person might get a "90" because they can't sleep and feel exhausted, while another gets a "90" because they feel worthless and have scary thoughts about hurting themselves. These are two very different stories, but the single number hides the difference. This study asks a big question: Is everyone with depression just one big, messy group, or are there hidden subgroups with their own unique "symptom recipes"? To find out, the researchers used a special kind of computer magic called "unsupervised learning." Imagine giving a robot a giant box of mixed-up Lego bricks and asking it to sort them into piles without telling it what a car or a house looks like. The robot has to find patterns on its own. The researchers also used a "shuffling trick" (a permutation null model) to make sure the robot wasn't just making up patterns where none existed, like seeing faces in clouds.

The Big Experiment

The researchers took data from a massive national health survey in the United States called NHANES, covering over 31,000 adults from 2007 to 2018. They decided to split the investigation into two parts. First, they wanted to see if people naturally fall into different "health types" based on things like their age, blood pressure, diet, and exercise. Crucially, they hid all the depression questions from the computer while it was sorting people. They wanted to see if the computer would find groups on its own, without knowing who was sad or who wasn't.

Part 1: The Health Types (The "Age" Clue)

When the computer tried to sort the 31,166 adults into health groups, it found a split, but it wasn't the exciting kind the researchers hoped for. The computer found two main groups, but they were mostly just "Younger People" and "Older People." The older group had more chronic diseases, higher blood sugar, and lower kidney function—basically, the usual health changes that come with getting older.

The computer was very consistent at making this split (it was 97.6% sure it was the same split every time it tried), and the groups were definitely different from random noise. However, the groups overlapped heavily, like two clouds that are mostly the same color. When the researchers checked how many people in each group had depression symptoms, the difference was small. The older group had about 10% of people with significant symptoms, while the younger group had about 7%. It was a real difference, but not a huge one. The main takeaway here is that if you just look at general health, you mostly just find age. The "health types" weren't distinct new categories; they were mostly just a restatement of the fact that older people have more health problems.

Part 2: The Symptom Recipes (The Real Discovery)

The second part of the study was where things got interesting. This time, the researchers looked only at the 7,861 adults who already reported having some depression symptoms. Instead of adding up their scores, they looked at the pattern of their answers. They asked: "If two people have the same total score, do they have the same mix of symptoms?"

To do this, they used a clever trick. They took each person's answers and divided them by their own total score. Imagine if you had a pizza and you cut it into slices. If you have a small pizza (low score) and a giant pizza (high score), this trick normalizes them so you can compare the shape of the slices, not the size of the pizza. This removed the "severity" (how bad they felt) and left only the "flavor" (what specific symptoms they had).

When they ran the computer on these "flavors," it found eight distinct subtypes of depression. These weren't just "mild," "moderate," and "severe." They were completely different recipes. Here is what the computer found:

Insomnia–Fatigue: People who mostly just couldn't sleep and felt tired. Only 4.8% of this group crossed the line into the "clinically significant" zone.
Anhedonic: People who mostly just couldn't feel joy.
Appetite–Somatic: People with changes in eating and physical feelings.
Depressed Mood: People who mostly just felt sad.
Concentration: People who mostly couldn't focus.
Self-Critical: People who mostly felt worthless.
Psychomotor: People who felt physically slowed down or agitated.
Affective–Suicidal: People who felt sad and had thoughts of self-harm. This group had the highest risk, with 58.4% crossing the clinical threshold.

Why This Matters

The most shocking part of the finding is that the "total score" is a terrible liar. The study showed that two groups could have almost the exact same average total score, but one group was mostly just tired and the other was mostly suicidal. For example, the "Anhedonic" group (joyless) and the "Appetite–Somatic" group (eating issues) had similar total scores, but their risk of being clinically depressed was totally different (13.4% vs 19.5%).

The researchers were very careful. They used a "shuffling test" to make sure the computer wasn't just hallucinating these eight groups. They scrambled the data so all the patterns were destroyed, and the computer failed to find these eight groups in the scrambled mess. This proved that the eight groups are real patterns in the data, not just a computer glitch.

The Bottom Line

This study suggests that treating depression as a single number is like trying to describe a whole orchestra by just saying "it's loud." Sometimes it's loud because of the drums (physical symptoms), and sometimes it's loud because of the violins (emotional thoughts). The "loudness" (total score) is the same, but the music is completely different.

The researchers found that while general health mostly just tells us about age, the way people experience depression is much more varied. There are at least eight different "flavors" of depression symptoms, and knowing which flavor someone has might be just as important as knowing how loud the music is. This doesn't mean we have solved the mystery of depression, but it does mean we need to stop looking at the single number and start listening to the specific notes each person is playing.

Technical Summary: Weak Health Phenotypes but Distinct Depressive-Symptom Subtypes among United States Adults

Problem Statement
Population studies of depression typically rely on two unexamined assumptions: first, that the population is homogeneous regarding the health characteristics accompanying depressive symptoms, and second, that a total symptom score (e.g., from the PHQ-9) adequately represents an individual's clinical picture. Conventional approaches often treat depression as a single dichotomy on a symptom scale. However, a total score of 12 could result from a combination of somatic symptoms (insomnia, fatigue) or cognitive-affective symptoms (anhedonia, suicidal ideation), which represent distinct clinical presentations with different correlates and treatment implications. Furthermore, existing literature on depression subtyping, primarily using Latent Class Analysis (LCA), has consistently found that identified subtypes are distinguished more by overall severity than by distinct symptom clusters, often failing to establish the existence of true symptomatic subtypes.

Methodology
This study utilized an unsupervised learning approach on six pooled cycles of the National Health and Nutrition Examination Survey (NHANES) from 2007 to 2018, comprising 31,166 adults for phenotype analysis and 7,861 adults (scoring $\ge$ 5 on PHQ-9) for symptom-subtype analysis.

Data Separation and Hygiene: Crucially, all depression-related variables (PHQ-9 items, totals, and derived outcomes) were programmatically excluded from the clustering feature sets to prevent circularity. A "firewall" assertion ensured no depression-derived variables entered the analysis.
Health Phenotype Clustering:
- Variables: 89 variables across nine domains (demographic, socioeconomic, anthropometric, lifestyle, sleep, nutrition, inflammation, chronic disease, clinical chemistry) were used.
- Dimensionality Reduction: Factor Analysis of Mixed Data (FAMD) was applied to handle the mix of continuous and categorical variables. 25 components were retained, explaining 63.4% of inertia.
- Clustering Algorithms: Three algorithms (K-means, Gaussian Mixture Models, and Ward hierarchical clustering) were tested. The number of clusters was determined by five internal validity criteria (Silhouette, Calinski–Harabasz, Davies–Bouldin, BIC, Gap statistic).
- Validation: Stability was assessed via bootstrap resampling. Crucially, a permutation null model was employed where variables were shuffled independently to destroy associations while preserving marginal distributions. This distinguished genuine structure from algorithmic artifacts.
Symptom Subtype Clustering:
- Raw vs. Person-Centred: Two approaches were tested on PHQ-9 item responses. The first used raw scores (recovering severity). The second used person-centred profiles, where item scores were divided by the individual's total score to remove severity and isolate symptom shape or pattern.
- Validation: Both approaches were tested against permutation nulls. The person-centred approach required a specific null because projecting data onto a simplex (dividing by row totals) can generate apparent structure even from independent variables.

Key Results

Health Phenotypes (Weak Structure):
- Five validity criteria disagreed on the optimal number of clusters (selecting 2, 2, 2, 6, and 10). The modal vote selected two clusters.
- While the two-cluster solution was highly reproducible (bootstrap adjusted Rand index 0.976) and significantly exceeded the permutation null (silhouette 0.1019 vs. null mean 0.0281), the clusters were poorly separated (one cluster had a negative silhouette score).
- The clusters largely represented an age and cardiometabolic axis: Cluster 1 was younger (mean age 38.3) with lower chronic disease burden, while Cluster 2 was older (mean age 61.2) with higher chronic disease burden and lower renal function.
- Depression Burden: The difference in depression burden between these two phenotypes was modest (9.98% vs. 6.88% prevalence; odds ratio 1.50), suggesting that broad health phenotypes do not strongly stratify depression risk beyond age and cardiometabolic status.
Depressive-Symptom Subtypes (Clear Structure):
- Raw Scores: Clustering raw PHQ-9 scores yielded two clusters that differed almost entirely by severity, replicating the diagnostic threshold without adding new clinical insight.
- Person-Centred Profiles: Removing severity revealed eight distinct subtypes that were stable under resampling and exceeded their permutation null by approximately 15 standard deviations.
- Subtype Characteristics: The eight subtypes ranged from 4.8% to 58.4% in the proportion of individuals crossing the clinical threshold (PHQ-9 $\ge$ $\geq$ 10).
  - Insomnia–Fatigue: Predominantly somatic, low cognitive-affective symptoms; only 4.8% crossed the threshold.
  - Affective–Suicidal: Predominantly cognitive-affective with suicidal ideation; 58.4% crossed the threshold.
  - Other Subtypes: Included Anhedonic, Appetite–Somatic, Depressed Mood, Concentration, Self-Critical, and Psychomotor.
- External Validity: The eight subtypes differed significantly on 13 of 15 external health variables (e.g., food insecurity, functional limitation, income-to-poverty ratio) that were not used in clustering. Somatic subtypes were associated with lower social adversity, while affective subtypes were associated with higher adversity.

Key Contributions

Methodological Rigor: The study demonstrates that unsupervised clustering claims must be validated against permutation nulls to distinguish genuine structure from algorithmic artifacts, particularly when using person-centred transformations that inherently create geometric structure.
Decoupling Severity and Pattern: By applying a person-centred transformation, the study successfully separated symptom pattern from symptom count, revealing that individuals with identical total scores can possess fundamentally different symptom architectures with vastly different clinical implications.
Empirical Evidence: The paper provides evidence that while broad health phenotypes in the US adult population are weakly separated and largely restatements of age, depressive symptom profiles exhibit a robust, multi-dimensional structure that total scores obscure.

Significance and Claims
The paper claims that symptom architecture carries information that total scores discard. The finding that subtypes with similar mean totals (e.g., Anhedonic vs. Appetite–Somatic) have vastly different rates of crossing the clinical threshold challenges the sufficiency of total scores in population research.

The authors are modest in their conclusions, noting that:

The identified subtypes are likely "regions of a symptom space" rather than discrete clinical entities.
The cross-sectional design precludes causal inference regarding whether these configurations precede or cause depressive symptoms.
The health phenotype result, while statistically real, is weakly separated and algorithm-dependent, suggesting that US adults do not partition into distinct health types based on these measurements.

The study concludes that meaningful heterogeneity in these data lies in symptom architecture, not in broad health phenotypes, and that future work should employ Latent Class Analysis and longitudinal follow-up to validate these findings.

Technical Summary: Weak Health Phenotypes but Distinct Depressive-Symptom Subtypes among United States Adults

More like this