NeuroQWERTY MIT-CSXPD: PD Keystroke Dataset

Updated 5 July 2026

The NeuroQWERTY MIT-CSXPD dataset is a clinically annotated free-text keystroke collection designed for Parkinson’s disease diagnosis and behavioral modeling.
It comprises 85 subjects, including 42 PD patients and 43 controls, with detailed keystroke timing data collected during naturalistic typing.
Researchers use derived temporal signals and minimal cleaning pipelines to explore transfer-learning and latent variable models linked to PD severity.

Searching arXiv for papers on NeuroQWERTY MIT-CSXPD and related keystroke-dynamics work. The NeuroQWERTY MIT-CSXPD dataset is a public, clinic-collected free-text typing dataset for research on Parkinson’s disease (PD) using keystroke dynamics. In the recent literature it is explicitly named “neuroQWERTY MIT-CSXPD” and treated as a single dataset rather than as separate “NeuroQWERTY” and “MIT-CSXPD” resources (Francesconi et al., 10 Oct 2025). The dataset contains 85 subjects, comprising 42 with PD and 43 healthy controls, and has been used both for PD diagnosis from keyboard dynamics and for severity-linked modeling based on raw keystroke telemetry (Francesconi et al., 10 Oct 2025). Its distinctive role in the field is as a clinically annotated benchmark built around naturalistic free-text typing, with raw key press and release timestamps sufficient to derive standard temporal keystroke signals and more structured behavioral models (Bondade, 24 Jun 2026).

1. Dataset identity and historical position

The dataset is described in recent work as a public neuroQWERTY MIT-CSXPD dataset available via PhysioNet, originally introduced by Giancardo et al. (2016) and subsequently reused in multiple strands of PD keyboard-dynamics research (Bondade, 24 Jun 2026). In the cross-dataset literature, it is designated DB2 and appears alongside TyPD, Tappy, and Online English as one of four public datasets for PD diagnosis via keyboard dynamics (Francesconi et al., 10 Oct 2025).

A central terminological point is that papers do not treat NeuroQWERTY and MIT-CSXPD as separate datasets. Rather, they use the combined name neuroQWERTY MIT-CSXPD for a single dataset (Francesconi et al., 10 Oct 2025). This matters because the literature on keyboard biomarkers contains several neighboring corpora with different collection settings, task structures, and modalities; conflating names can obscure whether a study is using the public clinic-collected touchscreen dataset or another typing corpus.

The dataset occupies an intermediate methodological position between tightly controlled motor assessments and unconstrained telemonitoring. It is clinic-collected, yet the task is free-text typing, described as participants typing spontaneously rather than reproducing fixed text (Francesconi et al., 10 Oct 2025). This combination has made it useful both as a standalone benchmark and as an adaptation domain in transfer-learning pipelines.

2. Cohort composition, session structure, and available metadata

Recent papers consistently report that the dataset contains 85 subjects, split into 42 PD and 43 healthy controls (Francesconi et al., 10 Oct 2025, Bondade, 24 Jun 2026). One study further resolves this into two sub-cohorts collected at different times:

Sub-cohort	Subjects	Sessions per subject
CS1	31	two recording sessions per subject
CS2	54	one session per subject

This yields the mixed session structure emphasized in the literature: CS1 is two-session, whereas CS2 is single-session, for a total of 85 subjects (Bondade, 24 Jun 2026). In the cross-dataset summary, the average number of sessions per subject is reported as 1.36 (Francesconi et al., 10 Oct 2025).

The dataset table in the cross-dataset PD study characterizes neuroQWERTY MIT-CSXPD as follows: #PD = 42, #HC = 43, Average signal length = $1492.59 \pm 644.76$ keystrokes per subject, Average number of sessions per subject = 1.36, Task = Free-text, and Context = Clinic (Francesconi et al., 10 Oct 2025). This is one of the clearest compact summaries of the dataset’s structure in current arXiv literature.

The metadata available to downstream users are more limited than the class labels might suggest. A 2026 study states that each PD subject has a same-day UPDRS-III motor score, as well as alternating finger tapping (afTap), single-key tapping (sTap), and the original study’s derived nqScore (Bondade, 24 Jun 2026). At the same time, that paper explicitly notes that the dataset provides no age or gender fields, preventing demographic confound adjustment. It also does not report medication state, disease duration, or stage, and these should therefore be treated as not reported in that source rather than assumed absent or present (Bondade, 24 Jun 2026).

A recurrent misconception is that the dataset is richly clinically annotated in the style of a full movement-disorders registry. The recent literature does not support that interpretation. What is documented is a clinically labeled typing dataset with PD/control labels and certain PD-specific motor measures, but without age and gender fields and without broader clinical stratification in the papers summarized here (Bondade, 24 Jun 2026).

3. Acquisition setting, task design, and raw data structure

The dataset is described as clinic-collected and based on a touchscreen keyboard during a free-text typing task (Francesconi et al., 10 Oct 2025). The wording “typing spontaneously as they would at home” is used to indicate that, although acquisition took place in clinic, the behavioral task was intended to preserve naturalistic typing rather than impose a fixed transcription paradigm (Francesconi et al., 10 Oct 2025).

Recent work is explicit that the dataset contains raw key press and release timestamps, rather than only summary features (Bondade, 24 Jun 2026). This distinction is methodologically important. It allows researchers to derive canonical temporal keystroke variables post hoc, test alternative preprocessing pipelines, and formulate sequence models at the level of transitions rather than relying on previously aggregated statistics.

The 2026 IRL study indicates that the following fields are explicitly available and used:

raw key press timestamps
raw key release timestamps
raw key characters
PD/control label
for PD subjects: UPDRS-III
for PD subjects: afTap
for PD subjects: sTap
for PD subjects: nqScore (Bondade, 24 Jun 2026)

That paper also states that key characters are used to infer hand alternation via a standard QWERTY touch-typing hand assignment, which implies the presence of symbolic key identity at the event level (Bondade, 24 Jun 2026).

The raw-event representation supports both standard keystroke-dynamics processing and more elaborate behavioral modeling. In particular, one can move from a stream of press and release events to inter-event timing variables, multivariate time-series windows, or transition-wise action models. This flexibility helps explain why the same dataset has been used in both classification pipelines and interpretable latent-variable studies.

4. Derived keystroke signals and preprocessing conventions

A dominant convention in recent work is to harmonize the dataset into four standard temporal signals:

$HT_i = R_i - P_i$

$FT_i = P_{i+1} - R_i$

$PP_i = P_{i+1} - P_i$

$RR_i = R_{i+1} - R_i$

where $P_i$ and $R_i$ denote press and release timestamps for key $i$ (Francesconi et al., 10 Oct 2025). In the cross-dataset study, all four signals for neuroQWERTY MIT-CSXPD are marked as Derived (D), meaning they were computed from available lower-level event timestamps rather than supplied as finished features (Francesconi et al., 10 Oct 2025).

The same paper adopts an intentionally minimal policy for data cleaning on this dataset. Because DB2 and DB3 are both free-text datasets, no data cleaning was applied to DB2, with the explicit rationale of preserving natural typing rhythm (Francesconi et al., 10 Oct 2025). Concretely, the paper states that for neuroQWERTY MIT-CSXPD there was:

no minimum typing-rate filter,
no outlier removal rule such as $FT > 3$ s,
no special denoising or correction (Francesconi et al., 10 Oct 2025).

By contrast, the 2026 IRL study applies a dataset-specific cleaning procedure following nqDataLoader.py and then adds an extra outlier filter. It removes rows with:

non-positive timestamps,
non-monotonic timestamps, or
hold times outside $[0,5)$ seconds,

which removes 3.1% of rows. It then removes flight-time outliers exceeding 3 seconds, accounting for 2.33% of the remaining rows (Bondade, 24 Jun 2026).

This divergence is not a contradiction so much as a reflection of different methodological priorities. The transfer-learning paper preserves free-text rhythm for cross-dataset diagnosis, whereas the IRL paper imposes stricter transition-level validity constraints because its reward model depends directly on local timing context. A plausible implication is that the dataset is robust enough to support multiple preprocessing philosophies, but that reported results are correspondingly sensitive to cleaning choices.

5. Principal analytic uses in the literature

The dataset has supported at least two distinct research programs in recent arXiv work: cross-dataset PD diagnosis and interpretable severity-linked behavioral modeling.

In the cross-dataset diagnosis study, neuroQWERTY MIT-CSXPD is DB2 and serves as the fine-tuning dataset within a staged transfer pipeline (Francesconi et al., 10 Oct 2025). The emphasized route is:

$HT_i = R_i - P_i$ 0

This role is specific: DB2 is not the large-scale pre-training source and not the final independent external test cohort. Instead, it functions as an adaptation bridge between larger source datasets and external validation (Francesconi et al., 10 Oct 2025). The rationale given is that DB2 has intermediate size and supports adaptation to a free-typing scenario.

The same study benchmarks eight deep-learning time-series architectures from the tsai library on windows of the four derived temporal signals: GRU, LSTM, GRU-FCN, LSTM-FCN, TCN, XCM, TSiT, and TSTPlus (Francesconi et al., 10 Oct 2025). Models operate on multivariate windows rather than on hand-engineered summary statistics alone, and evaluation is performed with subject-level separation using a leave-20%-subjects-out cross-validation scheme and patient-level aggregation of probabilities (Francesconi et al., 10 Oct 2025).

In contrast, the IRL study uses the same dataset primarily for severity-linked, interpretable behavioral modeling within the PD group, not for case-control classification (Bondade, 24 Jun 2026). Its unit of analysis is the keystroke transition, operationalized through flight time, and each transition is modeled as a discrete choice over flight time using a maximum-entropy discrete-choice model (Bondade, 24 Jun 2026). Flight times are discretized into $HT_i = R_i - P_i$ 1 bins via quintiles of the pooled flight-time distribution across all subjects, with each bin represented by its empirical mean flight time because the distribution is strongly right-skewed (Bondade, 24 Jun 2026).

The local context in that model is built from:

a 10-keystroke rolling mean of prior flight times, and
the immediately preceding flight time,

computed after bug correction using only prior keystrokes within the same recording session (Bondade, 24 Jun 2026). This use case illustrates that neuroQWERTY MIT-CSXPD is not merely a benchmark for classification accuracy; it is also a substrate for recoverable latent parameters tied to clinical severity.

6. Quantitative findings and benchmark role

The cross-dataset study reports the following dataset summary values for neuroQWERTY MIT-CSXPD: 42 PD, 43 HC, average signal length $HT_i = R_i - P_i$ 2 keystrokes per subject, 1.36 sessions per subject, free-text, clinic (Francesconi et al., 10 Oct 2025). Within the fine-tuning stage on DB2, transfer from DB4 markedly outperforms transfer from DB3. The DB4-to-DB2 fine-tuning results are reported as:

Model	AUC on DB2	F1 on DB2
GRU	75.20%	34.44%
LSTM	81.50%	54.56%
GRU-FCN	84.52%	88.87%
LSTM-FCN	88.94%	78.07%
TCN	88.04%	64.67%
XCM	82.64%	61.81%
TSiT	85.29%	64.44%
TSTPlus	79.15%	63.92%

These results indicate that neuroQWERTY MIT-CSXPD is an effective adaptation domain, especially when the source model is pre-trained on the larger Online English dataset (Francesconi et al., 10 Oct 2025). The paper highlights the external benefit of this intermediate adaptation: for example, TCN on the external TyPD cohort improves from 81.85% AUC / 62.68% F1 after DB4 pre-training only to 91.14% AUC / 79.39% F1 after DB2 fine-tuning (Francesconi et al., 10 Oct 2025). This establishes DB2 as a pivotal component in that transfer-learning framework.

The IRL paper’s headline finding is different in kind. On the neuroQWERTY MIT-CSXPD dataset, the recovered speed-preference weight correlates with UPDRS-III severity at

$HT_i = R_i - P_i$ 3

with the figure caption specifying that this is a Spearman correlation (Bondade, 24 Jun 2026). The result is further reported to replicate across the two sub-cohorts:

CS1: $HT_i = R_i - P_i$ 4
CS2: $HT_i = R_i - P_i$ 5 (Bondade, 24 Jun 2026)

The same study reports that raw typing speed alone explains

$HT_i = R_i - P_i$ 6

of UPDRS variance, while adding the recovered $HT_i = R_i - P_i$ 7 increases this to

$HT_i = R_i - P_i$ 8

with the increment significant at

$HT_i = R_i - P_i$ 9

(Bondade, 24 Jun 2026). This is a more specific claim than simply saying the dataset supports PD classification: it suggests that latent behavioral parameters inferred from raw keystroke timing carry explanatory value beyond mean speed alone.

The paper also reports negative results that are methodologically important. The recovered hand-alternation weight did not survive confound checks, and the consistency weight did not reach significance across tested settings (Bondade, 24 Jun 2026). This narrows the credible interpretation: on this dataset, the strongest robust signal in that framework is the recovered speed-preference parameter, not a broad suite of interpretable latent biomarkers.

7. Methodological caveats, validation issues, and relation to adjacent datasets

A defining feature of recent work on neuroQWERTY MIT-CSXPD is careful attention to leakage and context construction. The IRL paper documents two implementation bugs discovered during adversarial review:

session-boundary contamination, because rolling calculations initially did not reset between the two sessions of CS1 subjects;
rolling-window data leakage, because pandas.rolling().mean() initially included the current observation (Bondade, 24 Jun 2026).

After fixing session-boundary contamination, the headline correlation remained

$FT_i = P_{i+1} - R_i$ 0

and after fixing rolling-window leakage it became

$FT_i = P_{i+1} - R_i$ 1

(Bondade, 24 Jun 2026). The authors therefore argue that the main effect survives these corrections, while emphasizing that validation discipline is itself part of the scientific contribution.

Another caveat concerns feature identifiability. The same paper began with a four-parameter reward decomposition but found severe collinearity between two terms, with

$FT_i = P_{i+1} - R_i$ 2

in typical contexts, leading to an identifiable three-parameter model (Bondade, 24 Jun 2026). Even after the fix, residual collinearity between speed and consistency remained at

$FT_i = P_{i+1} - R_i$ 3

(Bondade, 24 Jun 2026). This indicates that the dataset supports interpretable modeling, but not without structural constraints and explicit collinearity checks.

In relation to adjacent typing datasets, neuroQWERTY MIT-CSXPD should be distinguished from both neural decoding corpora and non-clinical typing datasets. Brain2Qwerty is methodologically adjacent because it also involves QWERTY typing, keystroke timing, and motor interpretation, but it is a dataset of EEG/MEG aligned to typing in 35 healthy volunteers for brain-to-text decoding, not PD screening (Lévy et al., 18 Feb 2025). Likewise, emg2qwerty is a large public dataset of wrist sEMG during touch typing for sequence decoding and personalization, but it is a decoding and HCI benchmark rather than a clinically labeled PD biomarker dataset (Sivakumar et al., 2024). These comparisons help clarify the specific identity of neuroQWERTY MIT-CSXPD: it is a clinical keystroke-dynamics dataset centered on PD-related motor behavior, not a general-purpose neural or sEMG typing corpus.

A final misconception is that the dataset’s importance derives solely from high classification accuracy. The recent literature suggests a more nuanced view. One paper situates the dataset in relation to earlier neuroQWERTY results—AUC-ROC = 79% for Giancardo et al., AUC-ROC = 85% for Milne et al., and AUC-ROC = 85% for Roy et al.—but emphasizes that the newer contribution is cross-dataset transfer with external validation rather than a new single-dataset accuracy claim (Francesconi et al., 10 Oct 2025). Another uses the dataset not for headline diagnostic superiority, but for interpretable severity-linked latent-variable modeling and explicitly argues that the validation process itself matters in a literature with substantial heterogeneity (Bondade, 24 Jun 2026).

Taken together, these papers position the NeuroQWERTY MIT-CSXPD dataset as a compact but influential public benchmark: clinically grounded, behaviorally naturalistic, rich enough to support both discriminative and interpretable models, and methodologically demanding in ways that make leakage control, cohort-aware validation, and transparent preprocessing central to credible use (Francesconi et al., 10 Oct 2025, Bondade, 24 Jun 2026).