📄 medicine

From Health Checkups to Mental Health Referrals: Developing an AI-Based Depression Screening Model Using Routine Data in Older Adults Living Alone

This study developed and validated the DEPACS model, an explainable machine learning tool using routine health checkup data to effectively screen for depression in older adults living alone without relying on specific mental health assessments.

Original authors: Hyun Woo Jung, Jung Jae Lee

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hyun Woo Jung, Jung Jae Lee

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 a group of elderly people living by themselves. Often, they are struggling with sadness or depression, but no one knows it. They might not realize they are depressed, or they might feel too embarrassed to ask for help because of the stigma surrounding mental health. Usually, a doctor can only spot this if the patient admits, "I feel sad," or fills out a specific mental health questionnaire. But what if the patient never says a word about their feelings?

This research paper asks a bold question: Can a doctor tell if an older person is depressed just by looking at their regular physical checkup numbers, without ever asking them about their mood?

Here is how the researchers tried to solve this puzzle, explained simply:

The "Detective" vs. The "Medical Report"

Think of the researchers as detectives trying to solve a mystery. They had two tools:

The "Gold Standard" Detective (PHQ-9): This is a standard mental health questionnaire where people answer questions like "Do you feel down?" or "Do you have trouble sleeping?" The researchers used this first to see how well their computer programs (AI) could learn to spot depression when they had the obvious clues.
The "Physical Clues" Detective (DEPACS): This is the new tool they built. It only looks at things a doctor measures during a normal physical exam: heart rate, blood pressure, how well you can balance, how often you eat, how well you can smell or taste, and how big your waist is. It ignores all the "feeling sad" questions.

The Experiment

The team gathered data from 1,087 older adults living alone in South Korea. They fed this data into several different computer "brains" (algorithms like SVM, Random Forest, and XGBoost) to see which one could best guess who was depressed.

The Results:

The Gold Standard: When the computer used the mental health questionnaire (PHQ-9), it was a superstar. It was almost perfect, correctly identifying depression about 99% of the time. This proved the computer was smart enough to learn the patterns.
The Physical Clues (DEPACS): When the computer had to guess using only the physical checkup numbers (no mood questions allowed), it did much worse, but still did a "decent" job. It got it right about 74% of the time.

The researchers call this "fair" performance. It's not perfect, but it's good enough to say, "Hey, this person might be struggling, even though they didn't tell us."

The "Magic Lens" (SHAP Analysis)

To understand how the computer made its guesses, the researchers used a special "magic lens" called SHAP. This lens highlights which clues were the most important.

They found that the computer didn't just guess randomly; it looked for specific patterns that match what we already know about depression:

Sleep: People who slept poorly or woke up a lot were flagged as higher risk.
Movement: People who had trouble balancing or walking were flagged.
Heart Rate: A faster resting heart rate was a warning sign.
Habits: Eating fewer meals a day or exercising less was a red flag.
Senses: Trouble seeing, smelling, or tasting was also linked to depression.

Interestingly, the computer also noticed that larger waistlines were actually linked to a lower risk of depression in this group. The researchers suggest that for older adults, having a bit more weight might mean they are eating well and have enough money for food, whereas losing weight can sometimes be a sign of neglect or illness.

The "Noise" in the System

The researchers also tried a fun experiment: they removed the most "important" clues one by one to see if the computer got smarter or dumber.

When they removed clues like "how often you eat" or "how well you balance," the computer got dumber (its accuracy dropped). This confirmed these were real, useful clues.
However, when they removed head circumference (the size of the head), the computer actually got slightly smarter. This suggests that head size was acting like "static noise" on a radio—it looked important to the computer, but it was actually confusing the signal.

The Bottom Line

The paper concludes that while a computer can't replace a real mental health diagnosis, it can act as a safety net.

Imagine a doctor checking an elderly person's blood pressure and heart rate. If the numbers look a bit "off" (like a fast heart rate or poor balance), this AI model could whisper to the doctor, "This person might be depressed, even though they seem fine physically."

This is powerful because it catches people who are too shy or unaware to ask for help. It uses the routine data doctors already have to find the hidden cases of depression, offering a way to get these vulnerable people the care they need without them having to admit they are struggling first.

In short: You don't always need to ask "Are you sad?" to know someone might be. Sometimes, their heart rate, their balance, and their sleep habits tell the story for them.

Technical Summary: Developing an AI-Based Depression Screening Model Using Routine Data in Older Adults Living Alone

Problem Statement
Older adults living alone face an elevated risk of depression due to social isolation, physical decline, and economic hardship. However, this population remains largely undiagnosed due to limited symptom awareness, societal stigma, and insufficient referral systems. In South Korea, only 65% of patients with depression seek mental health care independently. Furthermore, somatic symptoms often mask underlying depression, leading to misattribution to physical illnesses. While digital phenotyping using wearables shows promise, it relies on non-routine biometric signals and lacks integration with standard clinical workflows. Consequently, there is a critical need for a screening mechanism that utilizes data already collected during routine general health checkups to enable primary care physicians to proactively identify at-risk individuals without requiring immediate psychiatric evaluation.

Methodology
This study employed a cross-sectional retrospective design using the Living Alone Older Adults (LAOA) dataset ( $n=1,087$ ), collected from the Intensive Managed Care for Elderly Living Alone (IMCELA) program in Cheonan, South Korea. The dataset included participants aged 65 and older who lived alone and had limited caregiving support.

Target Variable: Depression status was defined using the Short Form of the Geriatric Depression Scale (SGDS), where a score >10 indicated high risk.
Feature Engineering: The study developed the DEPACS (Depression Prediction based on Anthropometric and Clinical Signs) framework, incorporating 40 features across seven domains:
1. Socioeconomic status (gender, age, NHI status, education, marital status).
2. Vital signs (heart rate, respiratory rate).
3. Anthropometric measures (head, neck, and waist circumference).
4. Physical function (hearing, vision, olfactory, gustatory, gait, and balance performance).
5. Health behaviors (meal frequency, exercise duration, alcohol consumption, smoking exposure).
6. Chronic diseases (12 specific conditions and the Cumulative Illness Rating Scale).
7. Sleep quality (seven subcomponents of the Pittsburgh Sleep Quality Index).
Model Development: The dataset was split 80:20 into training and test sets. Five machine learning algorithms were evaluated: Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting, XGBoost, and K-Nearest Neighbors (KNN). Hyperparameter optimization was performed via grid search with 5-fold cross-validation.
Benchmarking: A control model using exclusively PHQ-9 items was developed to establish a performance ceiling.
Evaluation Metrics: Models were assessed using sensitivity, specificity, accuracy, F1 score, Brier score, Positive Predictive Value (PPV), Area Under the Precision-Recall Curve (AUPRC), and Area Under the Receiver Operating Characteristic Curve (AUROC).
Interpretability: SHapley Additive exPlanations (SHAP) analysis was applied to the best-performing models to identify feature contributions. A feature removal simulation was conducted to assess the impact of top predictors on model performance.

Key Results

Benchmark Performance: The PHQ-9 exclusive model demonstrated excellent performance, with most algorithms achieving AUROC and AUPRC scores exceeding 0.96 (SVM, Gradient Boosting, and XGBoost reached 0.99 AUROC).
DEPACS Model Performance: The DEPACS model, utilizing only routine health checkup data, achieved fair to acceptable predictive performance. The SVM algorithm yielded the best results with an AUROC of 0.74 and an AUPRC of 0.73. Other algorithms (Random Forest, Gradient Boosting, XGBoost) achieved AUROC scores between 0.71 and 0.73.
Feature Importance (SHAP Analysis):
- In the PHQ-9 model, "depressed mood" and "loss of interest" were the dominant predictors.
- In the DEPACS model, predictive power was distributed across multiple domains rather than concentrated in a single feature. Key predictors included sleep quality indices (subjective quality, latency, disturbances), waist and head circumference, balance function, heart rate, and meal frequency.
- Descriptive analysis confirmed that the depression group exhibited poorer sleep, lower meal frequency, impaired balance, higher heart rates, and higher rates of sensory dysfunction compared to the non-depressed group.
Feature Removal Simulation: Removing top SHAP-ranked features such as meal frequency, daytime dysfunction, and balance scores significantly decreased AUROC, confirming their utility. Conversely, removing head circumference slightly improved AUROC, suggesting it may act as a noise feature despite its high SHAP ranking.

Significance and Claims
The paper claims that the DEPACS framework offers a practical, explainable tool for primary care settings to identify depression risk in older adults living alone without relying on mental health-specific assessments.

Feasibility: The study demonstrates that meaningful prediction is feasible using only standard clinical data (vital signs, anthropometry, physical function, and sleep metrics), bypassing the stigma and barriers associated with direct psychiatric screening.
Clinical Utility: The model is positioned not as a replacement for clinical diagnosis but as a supplementary decision aid. It enables general practitioners to proactively recommend mental health evaluations to high-risk individuals who might otherwise avoid them.
Explainability: The use of SHAP analysis validates that the model's predictions align with established medical literature (e.g., the link between sleep disturbances, physical function decline, and depression), providing a transparent framework for clinical adoption.
Limitations: The authors acknowledge limitations regarding the single-center nature of the dataset (Cheonan, South Korea), which may affect generalizability. They also note that while the model performs well, its AUROC (0.74) is lower than specialized psychiatric tools, reinforcing its role as a screening rather than a diagnostic instrument.

Conclusion
This research validates the potential of integrating machine learning with routine health checkup data to bridge the treatment gap for depression in vulnerable older populations. By leveraging existing clinical workflows, the DEPACS model offers a pathway to early detection and referral, addressing the critical issue of underdiagnosis in socially isolated older adults.