PHQ-8: Depression Screening Tool

Updated 29 May 2026

PHQ-8 is a validated eight-item self-report tool that measures depression severity using DSM-aligned criteria.
It uses a 4-point Likert scale to score symptoms from 0 to 24 with established categorical cutoffs for minimal to severe depression.
Recent advances integrate PHQ-8 scoring with machine learning and multimodal analyses to improve interpretability and predictive accuracy.

The Patient Health Questionnaire-8 (PHQ-8) is a validated self-report instrument for the assessment of depressive symptomatology in both clinical and research contexts. It consists of eight items aligned with DSM criteria for major depressive disorder, designed to capture frequency and severity of core affective, cognitive, and somatic symptoms over the preceding two weeks. The PHQ-8 has achieved widespread adoption in epidemiological studies, digital health, and algorithmic psychiatric assessment due to its brevity, psychometric robustness, and suitability for both binary screening and dimensional severity estimation (Rosenman et al., 2024, Nerella et al., 17 Feb 2026, Zheng et al., 27 Jan 2025, Tang et al., 2024, Fara et al., 2022, Shi et al., 2024, Zhang et al., 2020).

1. Instrument Structure and Scoring

The PHQ-8 comprises eight items, each presented as a statement regarding depressive symptoms. Respondents rate symptom frequency using a 4-point Likert scale:

0 = Not at all
1 = Several days
2 = More than half the days
3 = Nearly every day

Let $x_i$ denote the score for the $i$ th item, $i \in \{1, \ldots, 8\}$ . The total score $S$ is the sum:

$S = \sum_{i=1}^{8} x_i$

The score ranges from 0 to 24. Standard cut points are widely used for categorical interpretation:

Range	Category
0–4	No/minimal depression
5–9	Mild depression
10–14	Moderate depression
15–19	Moderately severe
20–24	Severe depression

The PHQ-8 omits the suicidality item present in the PHQ-9, facilitating population-based screening where suicide assessment is impractical or raises ethical concerns (Tang et al., 2024, Shi et al., 2024, Zhang et al., 2020).

Standard Items

Little interest or pleasure in doing things (anhedonia)
Feeling down, depressed, or hopeless (depressed mood)
Trouble falling or staying asleep, or sleeping too much (sleep disturbance)
Feeling tired or having little energy (fatigue)
Poor appetite or overeating (appetite change)
Feeling bad about yourself, or that you are a failure or have let yourself or your family down (worthlessness)
Trouble concentrating on things, such as reading the newspaper or watching television (concentration)
Moving or speaking so slowly that other people could have noticed, or being so fidgety or restless that you have been moving around a lot more than usual (psychomotor change)

2. Applications in Computational and Multimodal Psychiatry

The PHQ-8 serves as a primary outcome in machine learning and artificial intelligence research targeting depression detection, severity regression, and symptom disaggregation. Recent work spans unimodal speech pipelines, multimodal fusion, and automated questionnaire completion with LLMs.

Speech and Cognitive Markers

Studies such as (Fara et al., 2022) isolate item-level associations between PHQ-8 subscales and biometric/behavioral features:

Speech features (e.g., loudness dynamics, spectral flux, syllable rate) predict cognitive and affective symptoms (anhedonia, worthlessness, concentration, etc.).
Cognitive markers (n-Back working memory metrics) preferentially index somatic and psychomotor subscales (sleep/fatigue, psychomotor change).

Fusion of these modalities modestly increases the area under the ROC curve (AUC ≈ 0.65 vs. 0.63 for best unimodal models).

Symptom-Guided Deep Learning

Contemporary architectures explicitly encode each PHQ-8 item as a trainable query in cross-attention mechanisms to align multimodal representations (primarily speech) with item content. The framework in (Nerella et al., 17 Feb 2026) computes, for participant $i$ and symptom $s$ , an attended representation $z_s$ over utterance embeddings $h_t$ :

$z_s = \sum_{t=1}^T a_{s,t} h_t$

where $i$ 0 is a softmax-weighted attention controlled by a per-symptom temperature parameter $i$ 1; $i$ 2 feeds a regression head predicting the item score $i$ 3. This approach yields interpretable, symptom-specific outputs and state-of-the-art root mean squared error ( $i$ 4) and mean absolute error ( $i$ 5), with per-item $i$ 6 as low as $i$ 7 for depressed mood.

Multimodal Explainable Models

The EMDRC framework (Zheng et al., 27 Jan 2025) leverages utterance-level PHQ-8 annotation (via LLMs), summary generation with transformers (LongT5), and cross-modal (text, audio, vision) fusion to produce symptom-explanatory summaries and scalar severity predictions. The system integrates utterance classification, symptom summary generation, and severity estimation losses, promoting interpretability and clinical relevance.

3. Algorithmic and LLM-Based PHQ-8 Scoring

LLMs are increasingly deployed to infer PHQ-8 scores from unstructured conversations or transcripts.

LLM Prompt Engineering and Completion

PHQ-8 items can be posed as prompts to LLMs, such as GPT-3.5 Turbo or custom-tuned GPT-4, with the model instructed to impersonate the interviewee and assign Likert scores (Rosenman et al., 2024). Variants of system prompts direct the model to analyze dialogue transcripts for explicit and implicit symptom cues, yielding a vector of integers (PHQ-8 responses) for input to downstream predictive models, such as Random Forest regressors.

Empirical performance is tracked using $i$ 8 and $i$ 9 against ground-truth PHQ-8 scores (when available): e.g., $i \in \{1, \ldots, 8\}$ 0, $i \in \{1, \ldots, 8\}$ 1 for LLM-prompted completion on DAIC-WoZ (Rosenman et al., 2024).

Chain-of-Thought (CoT) Reasoning

Embedding chain-of-thought prompts in LLM assessment (e.g., "Let’s think step by step: first, does the patient express loss of interest?...") significantly improves convergence to true self-reported PHQ-8 values. CoT prompting reduces per-item $i \in \{1, \ldots, 8\}$ 2 from ≈0.90 to ≈0.60 and overall $i \in \{1, \ldots, 8\}$ 3 from 2.4 to 1.6—a 33% improvement (Shi et al., 2024). This transparency mirrors clinical reasoning and enables item-level justifications, an asset in human-in-the-loop models.

Transparency-Driven Multi-Stage Scoring

"Psycho Analyst" (custom GPT-4), as detailed in (Tang et al., 2024), operationalizes a three-stage PHQ-8 computation: (1) holistic estimate, (2) itemized breakdown with evidence extraction, and (3) independent assessment/sanity check relative to reference scores. Reporting $i \in \{1, \ldots, 8\}$ 4 and $i \in \{1, \ldots, 8\}$ 5 after stage three, this pipeline demonstrates robust accuracy and justifiability.

4. Psychometric Validity and Empirical Properties

The PHQ-8 maintains high convergent validity with clinical diagnoses and physiological/behavioral markers of depression.

In digital phenotyping (Zhang et al., 2020), biweekly PHQ-8 self-report correlates with objectively measured sleep features from wearables (Fitbit), including sleep efficiency, insomnia, hypersomnia, and sleep-wake timing. Linear mixed-effects models show robust associations; for instance, a 10% increase in wake time at night ( $i \in \{1, \ldots, 8\}$ 6) predicts a 0.35-point higher PHQ-8 score. Instrumental stability across multiple sites and populations is documented, though site-specific variation highlights the need for context-sensitive modeling.

PHQ-8 item-level regression and classification demonstrate that symptom clusters can be differentiated via modality-specific signals (e.g., cognitive tasks vs. acoustic features) (Fara et al., 2022).

5. Interpretation, Limitations, and Future Directions

The PHQ-8's psychometric strengths must be tempered by recognized limitations and deployment considerations:

LLM-based item inference cannot always be validated against ground-truth respondent data when operating from unstructured interviews; hallucination and misinterpretation of context with numeric scales remain nontrivial risks (Rosenman et al., 2024).
Speech-only or text-only models capture cognitive/affective symptoms more reliably than internal/somatic states; the inclusion of multi-modal (audio, visual, behavioral) features is recommended where feasible (Nerella et al., 17 Feb 2026, Zheng et al., 27 Jan 2025).
Clinical cutoffs for significant depression (PHQ-8 ≥ 10) are effective for binary stratification but may not universally generalize across cultures and settings.
Algorithmic generalization requires thorough validation across diverse populations, languages, and interview formats.

Future work calls for more granular validation of LLM predictions against respondent answers, systematic inter-rater reliability studies, ethical scrutiny concerning privacy and bias, and the integration of adaptive, interactive interviewing agents leveraging real-time PHQ-8 reasoning (Tang et al., 2024, Shi et al., 2024).

6. Representative Tabular Summary: PHQ-8 Itemization

Item (No.)	Symptom Domain	Typical Wording
1	Anhedonia	Little interest or pleasure in doing things
2	Depressed mood	Feeling down, depressed, or hopeless
3	Sleep disturbance	Trouble falling/staying asleep, or sleeping too much
4	Fatigue	Feeling tired or having little energy
5	Appetite/weight	Poor appetite or overeating
6	Worthlessness/guilt	Feeling bad about yourself, failure, or letting down
7	Concentration	Trouble concentrating (reading, TV)
8	Psychomotor agitation/change	Moving/speaking slowly or restlessly

All items are scored on a 0–3 Likert scale (self-report), unless otherwise altered in specific computational setups (e.g., 1–5 in some LLM impersonation protocols) (Rosenman et al., 2024).

7. Conclusion

The PHQ-8 is central to both traditional and algorithmic paradigms for depression screening, anchoring symptom-level analyses and data-driven risk prediction. Its integration into LLMs, deep neural architectures, and multimodal fusion pipelines is enabling more item-specific, interpretable, and scalable psychiatric assessment, with robust grounding in empirical sleep, cognitive, and behavioral correlates (Rosenman et al., 2024, Nerella et al., 17 Feb 2026, Zheng et al., 27 Jan 2025, Tang et al., 2024, Fara et al., 2022, Shi et al., 2024, Zhang et al., 2020). Rigorous, context-aware deployment and continuous psychometric validation are essential for both research and clinical translation.