Multimodal longitudinal cognitive data for CNS R&D.
NLP-derived cognitive markers (TTR, MLU, coherence, idea density) validated against MoCA / MMSE — higher sensitivity to early change.
Daily data points vs. quarterly clinical visits. Track drug efficacy with weekly resolution instead of waiting 3–6 months for the next assessment.
Naturalistic data from home environments — not artificial lab settings. Ecological validity for FDA / EMA real-world evidence submissions.
Continuous voice-derived signal across the three CNS pillars.
Mood
Affective state inferred from vocal features, lexical markers and self-report — daily, baselined per subject.
Cognition
Composite Cognitive Index across recall, fluency, orientation, engagement — trended in 7-, 30- and 90-day windows.
Adherence
Two-way confirmed medication adherence, with missed-dose timestamps and downstream symptom correlations.
Why naturalistic data changes everything.
Existing digital cognitive tools rely on structured tests. Amigo captures data from natural daily conversations — no test anxiety, no practice effects, no patient burden.
| Dimension | Amigo | Structured test tools |
|---|---|---|
| Collection method | Free-form daily conversation | Prompted cognitive tasks (clock drawing, picture description, card games) |
| Frequency | Daily — every conversation is a data point | Weekly to quarterly — scheduled sessions |
| Patient burden | Zero — the conversation is the product (companionship) | Requires active engagement in a test protocol |
| Practice effects | None — no repeated test structure to learn | Scores improve with repeated exposure |
| Test anxiety | None — data collection embedded in an enjoyable daily routine | Performance anxiety can skew results |
| Ecological validity | High — real speech in home environment | Low to moderate — artificial test conditions |
| Data modalities | Linguistic + acoustic + behavioral + emotional | Typically 1–2 dimensions per tool |
| Adherence | High — seniors want to talk | Drop-off over time — test fatigue |
| Longitudinal depth | Hundreds of sessions per subject over months | Tens of sessions at best |
| Cost per data point | Marginal — no clinical staff, no device provisioning | Higher — supervised administration or dedicated devices |
“Structured test tools” refers to digital cognitive assessment platforms based on prompted tests (clock drawing, picture description, reaction-time tasks) — a category that requires active patient engagement.
From natural voice to research-ready structured datasets.
Every step is engineered and operated in-house, with explicit versioning at each stage and a full audit trail of how each data point was produced.
31 structured variables per session.
Across cognitive, vocal, behavioral, emotional and clinical dimensions. Schema versioned, data-dictionary delivered with every export.
| Field | Type | Example | Category |
|---|---|---|---|
| subject_id | string | anon_8f3a2c | ID |
| session_date | date | 2026-02-28 | ID |
| call_duration_seconds | int | 540 | Session |
| word_count | int | 312 | Cognitive |
| type_token_ratio | float | 0.68 | Cognitive |
| mean_utterance_length | float | 8.4 | Cognitive |
| words_per_minute | float | 98 | Cognitive |
| repetition_rate | float | 0.05 | Cognitive |
| coherence_score | int (1-10) | 8 | Cognitive-AI |
| idea_density_score | int (1-10) | 7 | Cognitive-AI |
| word_finding_score | int (1-10) | 9 | Cognitive-AI |
| composite_score | int (0-1000) | 742 | Cognitive |
| participation_ratio | float | 0.45 | Behavioral |
| turn_count | int | 24 | Behavioral |
| engagement_quality | enum | good | Behavioral |
| mood_primary | enum (5) | positive | Mood |
| mood_intensity | int (0-10) | 7 | Mood |
| alert_triggered | bool | false | Alert |
| alert_nature | enum | null | Alert |
| alert_severity | enum | null | Alert |
| f0_mean_hz | float | 185.3 | Voice |
| f0_std_hz | float | 28.7 | Voice |
| jitter_percent | float | 1.12 | Voice |
| shimmer_percent | float | 3.45 | Voice |
| hnr_db | float | 18.2 | Voice |
| speech_rate | float | 3.8 | Voice |
| pause_rate | float | 0.22 | Voice |
| mean_pause_duration_ms | int | 680 | Voice |
| composite_z_score | float | +0.3 | Baseline |
| baseline_sessions | int | 12 | Baseline |
| trend | enum | stable | Baseline |
Cohort breakdown
Distribution of clinical profiles across the dataset
Metric distribution
Sample histogram of AI-assessed coherence scores
Voice frequency evolution
F0 mean (Hz) over 24 weeks — one subject
What every conversation produces.
Each call generates a multi-dimensional structured snapshot. Versioned, time-stamped, and lineage-traceable.
NLP cognitive markers
Quantitative linguistic metrics from speech patterns via Amigo's deterministic NLP pipeline.
AI cognitive scores
Higher-order language function scored by the Amigo Cognitive Engine, calibrated against clinical scales.
Voice & acoustic biomarkers
Prosodic and spectral features extracted from raw audio. Validated markers for depression, Parkinson's and cognitive decline.
Behavioral metrics
Engagement, participation and interaction dynamics — revealing behavioral patterns over time.
Mood & affect
5-class mood classification with continuous intensity scoring and session-level emotional context.
Clinical alerts
Structured detection of depression markers, cognitive confusion, physical complaints and safety signals.
Longitudinal baselines
Intra-individual z-score baselines enabling personal trajectory tracking and change detection.
One conversation. One structured record.
Each record represents one session. Available in JSON, CSV or Parquet — with a versioned data dictionary and audit lineage attached.
{
"subject_id": "anon_8f3a2c",
"age_band": "75-79",
"session_date": "2026-02-28",
"call_duration_seconds": 540,
"cognitive": {
"word_count": 312,
"type_token_ratio": 0.68,
"mean_utterance_length": 8.4,
"repetition_rate": 0.05,
"words_per_minute": 98,
"coherence_score": 8,
"idea_density_score": 7,
"word_finding_score": 9,
"composite_score": 742
},
"voice": {
"f0_mean_hz": 185.3,
"f0_std_hz": 28.7,
"jitter_percent": 1.12,
"shimmer_percent": 3.45,
"hnr_db": 18.2,
"speech_rate": 3.8,
"pause_rate": 0.22,
"mean_pause_duration_ms": 680
},
"behavioral": {
"participation_ratio": 0.45,
"turn_count": 24,
"engagement_quality": "good",
"user_speech_duration_s": 245
},
"mood": {
"primary": "positive",
"intensity": 7
},
"alert": {
"triggered": false,
"nature": null,
"severity": null
},
"baseline_delta": {
"composite_z_score": +0.3,
"sessions_in_baseline": 12,
"trend": "stable"
}
}Grounded in peer-reviewed research.
Preliminary internal correlations with established clinical scales. External multi-site validation in progress.
Metric ↔ clinical scale correlations
Preliminary · internal validation cohort (N = 120)
| Amigo metric | Scale | r | Direction |
|---|---|---|---|
| Composite Score | MoCA | 0.72 | ↑↑ |
| Type-Token Ratio | MMSE | 0.65 | ↑↑ |
| Idea Density | MoCA | 0.68 | ↑↑ |
| Word Finding Score | BNT | 0.71 | ↑↑ |
| Repetition Rate | MMSE | -0.58 | ↑↓ |
| Pause Rate | MoCA | -0.54 | ↑↓ |
| F0 Std (Hz) | GDS | -0.49 | ↑↓ |
| Mood Intensity | GDS | -0.61 | ↑↓ |
| Jitter | UPDRS-III | 0.52 | ↑↑ |
| Speech Rate | MoCA | 0.47 | ↑↑ |
Validation methodology
- 1
Paired assessment: Subjects completed standard clinical scales (MoCA, MMSE, GDS, BNT) within 48h of Amigo conversation sessions.
- 2
Pearson correlation: Each metric was correlated with the clinically closest scale at subject level, controlling for age and education.
- 3
Longitudinal sensitivity: Change scores over 6 months vs. clinical re-assessment — Amigo composite detected decline 4–8 weeks earlier than quarterly MoCA.
Grounded in published research
- →
Type-Token Ratio decline as MCI marker — Bucks et al., 2000; Forbes-McKay & Venneri, 2005
- →
Idea density predicting AD onset — Snowdon et al. (Nun Study), 1996
- →
Pause patterns and speech rate in cognitive decline — Hoffmann et al., 2010; Roark et al., 2011
- →
F0 and jitter as depression biomarkers — Cummins et al., 2015; Low et al., 2011
- →
Vocal shimmer / HNR in Parkinson's disease — Tsanas et al., 2012; Rusz et al., 2011
- →
Spontaneous speech analysis for dementia screening — Fraser et al., 2016; Luz et al., 2020
Seeking academic and pharma partners for multi-site validation studies.
Where the data is being used.
Across neurodegenerative disease, geriatric psychiatry, clinical drug development and real-world evidence.
Alzheimer's & dementia early detection
Train models on subtle linguistic markers (declining TTR, increased repetition, reduced idea density) that precede clinical diagnosis by months or years. Longitudinal baselines enable intra-subject change detection.
Clinical trial digital endpoints
Continuous cognitive snapshots as digital endpoints for Phase II–III trials. Weekly resolution vs. quarterly MMSE — higher sensitivity to change, lower patient burden, remote monitoring.
Geriatric depression & social isolation
Longitudinal mood trajectories, engagement decline patterns and alert frequency data for studying depression onset, social isolation and intervention effectiveness.
Real-world evidence for regulatory
Naturalistic, continuous data from real home settings. Ideal for FDA / EMA post-market surveillance, label expansion and real-world evidence packages.
Voice & acoustic biomarker research
Track F0 decline, jitter / shimmer evolution and pause pattern changes over months. Validated vocal markers for depression screening, Parkinson's monitoring, early cognitive change.
Speech & linguistic biomarker validation
Cross-validate computational linguistic and acoustic features against established clinical cognitive scales. Paired metric + score data enables robust multi-modal biomarker validation.
Built for the populations CNS trials reach hardest.
Voice-first acquisition removes most of the friction that limits digital endpoint capture in older or cognitively vulnerable populations.
Geriatric depression
Daily mood signal between site visits with weak-signal alerting.
MCI & early-stage Alzheimer's
Longitudinal cognitive markers without repeated formal testing.
Parkinson's & motor disorders
Speech-based features and adherence reporting across home setting.
Schizophrenia & bipolar
Daily affect and engagement signals between clinical contacts.
Pain & long-COVID
Patient-reported symptom and impact tracking via natural conversation.
Caregiver / dyad studies
Parallel signal capture from family informants where appropriate.
Built to slot into the research stack you already operate.
Voice-only acquisition
No wearables, no apps, no participant burden. The cohort answers a phone — that's it.
API-first data flow
Per-participant time-series export to your data lake. Compatible with REDCap and standard CDISC mappings on request.
Audit-grade record
Every call, transcript, score, alert and access event is timestamped, immutable and exportable.
Regulatory posture
HIPAA-aligned controls, role-based access, encryption in transit and at rest, BAA available, 21 CFR Part 11 roadmap on request.
Designed for the assumptions sponsors and IRBs already hold.
Flexible data access tailored to your program.
Static dataset
One-time historical export. Ideal for exploratory analysis, model training and feasibility studies.
- Full historical cohort
- CSV / Parquet / JSON
- Data dictionary included
Live data feed
Ongoing API access with daily increments. For clinical monitoring and adaptive trial designs.
- Real-time API access
- Daily incremental updates
- Webhook notifications
- SLA-backed uptime
Research partnership
Co-designed protocol with custom metrics, targeted cohort criteria, and joint publication rights.
- Custom NLP pipelines
- Targeted cohort selection
- Joint publication
- Dedicated data engineer
Get a sample dataset with 1,000 anonymized sessions.
No commitment. Evaluate the data before any licensing discussion. We'll come back with a fit assessment for your protocol within five business days.