VeteranAD: Driving & Health Analytics

• VeteranAD is a homonymous label defined in two distinct fields: an end-to-end autonomous driving framework and veteran health analytics.
• In autonomous driving, it couples multi-view perception with planning using anchor-conditioned, autoregressive decoding for improved performance.
• In veteran health, it underpins early Alzheimer’s screening and readmission risk prediction through analysis of EHR and eye-tracking data.

Searching arXiv for the provided identifiers and topic usage to ground the article in current records. arxiv_search(query="VeteranAD OR \"Perception in Plan\" OR (Zhang et al., 15 Aug 2025) OR (Li et al., 2023) OR (Sriram et al., 2023) OR (Ajorlou et al., 2014)", max_results=10, sort_by="relevance") Searching arXiv for relevant papers. {"query":"VeteranAD OR \"Perception in Plan\" OR (Zhang et al., 15 Aug 2025) OR (Li et al., 2023) OR (Sriram et al., 2023) OR (Ajorlou et al., 2014)","max_results":10,"sort_by":"relevance"} VeteranAD denotes distinct research constructs in contemporary technical literature. One usage names a camera-only end-to-end autonomous driving framework that couples perception directly to evolving planning intent through a "perception-in-plan" design (Zhang et al., 15 Aug 2025). A separate usage applies the same label to veteran-centered clinical analytics, especially early Alzheimer’s disease screening from longitudinal Veterans Health Administration notes, with related extensions involving eye-tracking classification and Department of Veterans Affairs readmission-risk modeling (Li et al., 2023, Sriram et al., 2023, Ajorlou et al., 2014). The coexistence of these usages suggests a homonymous label rather than a single research lineage.

1. Terminological scope and domain separation

The term spans at least two unrelated methodological domains. In autonomous driving, VeteranAD is an explicitly named architecture whose central contribution is coupled perception and planning under multi-mode anchored trajectory priors and autoregressive rollout (Zhang et al., 15 Aug 2025). In veteran healthcare analytics, the label is used for workflows centered on Alzheimer’s disease risk screening from longitudinal records and for operational risk models in VA populations (Li et al., 2023, Ajorlou et al., 2014).

Usage of VeteranAD	Domain	Core substrate
VeteranAD	End-to-end autonomous driving	Multi-view camera inputs, BEV features, agent features, anchored trajectories
VeteranAD screening pipeline	Early Alzheimer’s disease prediction in veterans	Longitudinal VHA EHR notes, keyword-age pairs
VeteranAD ET augmentation	Alzheimer’s classification support	Raw eye-tracking sequences and scanpath images
VeteranAD readmission workflow	VA risk stratification	Recurrent-event inpatient survival data

This separation matters because the same name can otherwise obscure incompatible assumptions, evaluation criteria, and deployment environments. In the driving literature, the relevant quantities are PDMS, DAC, TTC, EP, and trajectory error. In the clinical literature, the relevant quantities are ROC AUC, PR AUC, calibration, survival hazard, and subgroup robustness. A common misconception is therefore to treat VeteranAD as a unified platform; the available record instead indicates domain-specific constructs that share only a label.

2. VeteranAD as a perception-in-plan autonomous driving framework

In autonomous driving, VeteranAD advances mainstream end-to-end systems that chain perception and planning sequentially inside a differentiable architecture. The motivating claim is that planning-oriented optimization alone is insufficient: gradients from planning losses may update perception, but perception remains a largely generic scene parsing stage unless it is explicitly conditioned on the immediate plan (Zhang et al., 15 Aug 2025). VeteranAD addresses this with a perception-in-plan framework in which planning priors and perception are coupled throughout rollout.

The input is multi-view cameras $$I \in \mathbb{R}^{N_{\text{img}} \times 3 \times H \times W}$$, using three cameras—front-left, front, and front-right—resized to $768 \times 432$. A ResNet-34 image encoder produces image features $F_{\text{img}}$, and an LSS-based ImageToBEV module forms $F_{\text{BEV}}$:

$$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$

Agent features are obtained by Transformer interaction with BEV features:

$$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$

Planning intent is represented by multi-mode trajectory queries $Q_{\text{traj}} \in \mathbb{R}^{M \times C}$, with $M=20$ planning modes initialized from anchored trajectories clustered by K-Means over ground-truth ego trajectories. Each anchored trajectory yields guiding points $P_t \in \mathbb{R}^{M \times 3}$, which act simultaneously as perception reference points and coarse planning priors. This dual use is the core coupling interface.

Planning-Aware Holistic Perception position-guides cross-attention over image and BEV features:

$$Q_{\text{traj}} = \text{CrossAttn}(Q = Q_{\text{traj}}, K, V = F_{\text{img}}),$$

$768 \times 432$0

The guiding points are projected into camera and BEV coordinates, focusing sampling along candidate paths in the style of DETR3D and SparseBEV. Agent interaction is additionally distance-aware. Relative distances between ego guiding points and detected agents are encoded into $768 \times 432$1, concatenated to yield

$768 \times 432$2

followed by agent cross-attention:

$768 \times 432$3

This implements soft ROI selection over map elements, drivable area, traffic lights, road geometry, and nearby actors along the intended path.

Trajectory generation is localized and autoregressive. At each time step, an MLP decoder predicts a local offset $768 \times 432$4 that refines the anchor:

$768 \times 432$5

Rollout proceeds from near to far future, yielding

$768 \times 432$6

and final mode scores $768 \times 432$7. Motion-Aware Layer Normalization conditions $768 \times 432$8 from time $768 \times 432$9 to $F_{\text{img}}$0 on the current guiding point, preserving temporal dynamics. The top-scoring mode at the final step is selected for execution, while the remaining modes provide multi-modal alternatives for training and evaluation.

The overall training objective is

$F_{\text{img}}$1

Here $F_{\text{img}}$2 is cross-entropy on BEV segmentation, $F_{\text{img}}$3 combines L1 box regression and BCE labels, $F_{\text{img}}$4 is L1 waypoint regression, and $F_{\text{img}}$5 is focal loss over the multi-mode scores. The weights are $F_{\text{img}}$6 on NAVSIM and all $F_{\text{img}}$7 on Bench2Drive.

3. Benchmark performance, ablations, and failure modes in driving

VeteranAD was evaluated on NAVSIM, Bench2Drive, and as an integration atop VAD on nuScenes, with results indicating that perception-in-plan coupling improves both open-loop and closed-loop planning metrics (Zhang et al., 15 Aug 2025). On NAVSIM navtest, the camera-only ResNet-34 variant reports NC 99.1, DAC 98.3, TTC 96.1, Comfort 100, EP 83.1, and PDMS 90.2. The same report states that it outperforms DiffusionDrive at PDMS 88.1, WoTE at 88.3, DRAMA at 85.5, and Transfuser, UniAD, Para-Drive, VADv2, and Hydra-MDP, despite using fewer modalities.

On Bench2Drive, the open-loop average L2 displacement error is 0.60 m, compared with 0.62 m for DriveTransformer. In closed loop, VeteranAD achieves Driving Score 64.22 and Success Rate 33.85%, compared with DriveTransformer at DS 63.46 and SR 35.01, DriveAdapter at DS 64.12 and SR 33.08, and ThinkTwice at DS 62.44 and SR 31.23. On nuScenes open-loop planning when integrated atop VAD, the average L2 decreases from 0.72 m to 0.62 m and the average collision rate decreases from 0.22% to 0.16%, while runtime changes from 4.5 FPS for VAD to 4.3 FPS.

Ablation studies isolate the contributions of the coupling mechanism. On NAVSIM, using only Localized AR Planning without position-guided perception yields PDMS 87.5, using only Planning-Aware Holistic Perception without anchor-guided offsets yields PDMS 88.4, and the full coupling yields PDMS 90.2. Removing any of the image, BEV, or agent attention pathways degrades performance, with the largest drop resulting from removal of BEV attention; the full configuration scores 90.2 PDMS, whereas partial variants fall in the 89.1–89.7 range. Autoregressive decoding also outperforms one-shot non-autoregressive decoding: NAR gives PDMS 88.6, DAC 97.1, TTC 94.9, and EP 82.1, while AR gives PDMS 90.2, DAC 98.3, TTC 96.1, and EP 83.1.

The computational profile is moderate rather than minimal. Training is reported at approximately 8 hours on 8 NVIDIA GeForce RTX 3090s with total batch size 32 for 16 epochs, compared with approximately 9 hours for DiffusionDrive under its official protocol. Inference latency is 22.3 ms per frame, compared with 18.4 ms for DiffusionDrive.

The published limitations are also specific. Closed-loop robustness remains challenging, with Success Rate still modest at 33.85% on Bench2Drive. Reported failure cases include wrong-road selection in NAVSIM, imprecise turning paths despite correct high-level intent, and off-road behavior during closed-loop turning. Proposed future directions include reinforcement learning or safety constraints in the end-to-end loop, explicit navigation signals and traffic rule priors, and broader prior coverage for rare maneuvers and occlusions off the anchored path.

4. VeteranAD as longitudinal EHR-based Alzheimer’s risk screening in veterans

A separate clinical usage applies VeteranAD to early Alzheimer’s disease prediction in U.S. military veterans using longitudinal VHA electronic health records (Li et al., 2023). The underlying study used de-identified VHA Corporate Data Warehouse records from January 1, 2004 to October 1, 2021 in a case-control design. Incident AD cases required a first ICD-based diagnosis on or after 1/1/2016, with additional clinic- and provider-type restrictions intended to strengthen case ascertainment. Cases and controls were matched 1:9 with replacement by age, sex, and clinical utilization. The final cohort contained 16,701 cases and 39,097 matched controls.

The predictive representation is based on 131 AD-relevant keywords curated by domain experts from the literature, spanning cognition, physiology or behavior, mood, and cognitive or diagnostic testing. Examples include “concentration,” “speaking,” “language,” “memory,” “MMSE,” “Mini-Cog,” “affect,” “judgement,” “lost,” and “sleep.” For each patient, keyword occurrences are aligned to age, producing keyword-age pairs $F_{\text{img}}$8. Across the cohort, 8,751 unique keyword-age pairs were identified. Term frequency and inverse document frequency analysis was used to retain the most frequent 1,000 keyword-age pairs as predictors, and the model input consists of frequencies of these selected pairs within the observation window.

Prediction horizons are implemented through a clean window of length $F_{\text{img}}$9 years immediately preceding the index date, so that predictions at horizon $F_{\text{BEV}}$0 use only data earlier than $F_{\text{BEV}}$1. Patients with fewer than five years of EHR history in the observation window were excluded. Four machine-learning models were evaluated: logistic regression, support vector machine, AdaBoost, and random forests. Two validation settings were used: Setting I, an 80/20 random split, and Setting II, an “external-like” holdout of 10 VHA stations.

The strongest reported results come from logistic regression. At the 10-year horizon in Setting I, with test set size 4,076 and 1,112 AD cases, logistic regression achieves accuracy 0.96, AD-class precision/recall/F1 of 1.00/0.87/0.93, control-class precision/recall/F1 of 0.95/1.00/0.98, PR AUC 0.990, and ROC AUC 0.997. In Setting II, with test set size 1,670, logistic regression again achieves accuracy 0.96, AD-class precision/recall/F1 of 1.00/0.86/0.93, control-class precision/recall/F1 of 0.95/1.00/0.97, PR AUC 0.998, and ROC AUC 0.991. Calibration is described as excellent, with Hosmer-Lemeshow goodness-of-fit $F_{\text{BEV}}$2 and calibration plots close to the $F_{\text{BEV}}$3 line.

The logistic model uses the standard predictive probability

$F_{\text{BEV}}$4

The study reports that ICD codes alone perform poorly as a baseline, with ROC AUC approximately 0.536 at the 1-year horizon and approximately 0.500 at the 10-year horizon.

Longitudinal signal analysis is central to the argument. Cases show AD-related keyword counts rising from around 10 mentions per year to over 40 as diagnosis approaches, whereas controls remain approximately flat at 10 per year. Divergence begins about 14 years before diagnosis and persists after normalization by note counts, when restricted to primary care notes, and across diagnosis-age bands. SHAP analysis identifies “MMSE,” “concentration,” “speaking,” “language,” “affect,” and “judgement” among the most influential predictors, with age-varying prominence of testing, memory, and speech-language groups.

The principal limitation is subgroup heterogeneity. Performance is substantially lower for patients younger than 65, with ROC AUC approximately 0.746 and accuracy approximately 0.76 for 10-year prediction in that subgroup. The paper also notes context-insensitive keyword matching, no explicit handling of negation or synonymy, VHA-specific demographics, and possible misdiagnosis overlap with MCI, MDD, and non-AD dementias.

5. Eye-tracking deep learning as an auxiliary AD-screening modality

An additional line of work relevant to veteran-focused AD screening uses deep learning on raw eye-tracking data rather than EHR notes (Sriram et al., 2023). The model, VTNet, was evaluated on data originally collected by Jang et al. and later expanded. The final dataset used in the study contains 69 patients and 75 controls. Patients were from a specialized memory clinic and included Alzheimer’s disease or early mild cognitive impairment potentially leading to AD; controls were from the community and were age- and sex-matched. Eye movements were recorded with a Tobii Pro X3-120 at 120 Hz, and each raw sample is a 6D vector containing gaze position, binocular distance-to-screen, and pupil size.

Three tasks were studied: Pupil Calibration, Picture Description, and Reading. Sequence lengths are long for the latter two tasks, with overall averages above 8,000 timesteps for picture description and reading and maxima above 26,000. To manage this, the study used cyclical splitting, in which each original sequence is divided into four interleaved sequences, reducing length by a factor of four while augmenting the number of data points by the same factor. For the longest tasks, post-split maximum sequence length cutoffs of 1,000 and 2,000 timesteps yielded VTNet_1000 and VTNet_2000 variants.

VTNet combines a single-layer GRU with 256 hidden units and a two-layer CNN operating on a scanpath image showing where fixations occurred and the transitions between them. The GRU branch processes raw 6D sequences, while the CNN branch yields a 50-dimensional representation. Their outputs are concatenated into a 306-dimensional fused representation, followed by a fully connected network with SoftMax output. A lightweight self-attention layer was also placed before the GRU, using PyTorch Multi-Head Attention with dimension 6 and one head.

The GRU update equations are reported as

$F_{\text{BEV}}$5

$F_{\text{BEV}}$6

$F_{\text{BEV}}$7

$F_{\text{BEV}}$8

The CNN branch uses a generic convolution

$F_{\text{BEV}}$9

Performance is task-dependent and improves materially with attention. For Pupil Calibration, VTNet_full_att achieves AUC 0.78, sensitivity 0.71, and specificity 0.75, compared with a baseline AUC of 0.71. For Picture Description, VTNet_1000_att achieves AUC 0.76, sensitivity 0.70, and specificity 0.73, compared with a baseline AUC of 0.75. For Reading, VTNet_full_att achieves AUC 0.78, sensitivity 0.70, and specificity 0.80, compared with a baseline AUC of 0.74. Statistical analysis used one-way MANOVA with Tukey’s HSD post-hoc testing, and evaluation used 10 runs of 10-fold user-level stratified cross-validation.

This work is not veteran-specific in cohort construction, but its discussion explicitly considers veteran deployment. The proposed relevance is that fixation stability, scanpath organization, and oculomotor reading behavior are likely affected in AD independent of setting, while an end-to-end approach may reduce reliance on task-specific engineered features in populations with PTSD, TBI, visual changes, or medication effects. A plausible implication is that VTNet functions as a modality extension to a VeteranAD screening concept rather than as a veteran-validated standalone system.

6. VeteranAD in VA readmission-risk modeling and broader interpretation

A further clinical usage associates VeteranAD with VA readmission-risk prediction using hierarchical frailty models (Ajorlou et al., 2014). The target outcome is 30-day hospital readmission judged to be caused primarily by patient non-compliance with medication and post-discharge instructions. The study used data from four Michigan VA Medical Centers—Ann Arbor, Battle Creek, Detroit, and Saginaw—covering the first halves of the years 2008 through 2012. The sample includes 6,685 admissions from 5,180 unique inpatients, with 467 readmissions in the observation window, corresponding to 9.02%.

The methodological choice is recurrent-event survival analysis with patient-level frailty. The hazard definition is

$$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$0

The proportional hazards form is

$$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$1

and the estimated frailty model is

$$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$2

Here $$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$3 indexes patients and $$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$4 indexes successive discharges or readmission opportunities. The baseline hazard is Weibull, with estimated $$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$5, interpreted in the paper as effectively exponential over the 30-day window.

Covariates retained after feature engineering and stepwise selection include admission source, patient sequence, $$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$6, the interactions sequence $$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$7 admission source and sequence $$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$8, and marital status $$F_{\text{img}} = \text{ImageEncoder}(I), \qquad F_{\text{BEV}} = \text{ImageToBEV}(F_{\text{img}}).$$9 user enrollment. Estimation used exact likelihood with adaptive Gauss-Hermite quadrature and dual quasi-Newton optimization. The paper reports strong random-effect heterogeneity, with frailty standard deviation $$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$0.

Selected parameter estimates illustrate the structure of risk. The intercept is $$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$1. Admission source has coefficient $$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$2, sequence $$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$3 admission source has coefficient $$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$4, sequence $$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$5 has coefficient $$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$6, married $$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$7 enrolled has coefficient $$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$8, and previously married $$F_{\text{agent}} = \text{Transformer}(Q = F_{\text{agent}}, K, V = F_{\text{BEV}}).$$9 enrolled has coefficient $Q_{\text{traj}} \in \mathbb{R}^{M \times C}$0. The joint delta-method F-test for the marital-status interaction gives $Q_{\text{traj}} \in \mathbb{R}^{M \times C}$1. The reported generalized $Q_{\text{traj}} \in \mathbb{R}^{M \times C}$2 is 79.68%.

Predicted hazards are highly right-skewed. Reported summary values include minimum approximately $Q_{\text{traj}} \in \mathbb{R}^{M \times C}$3, median approximately 0.06, mean approximately 0.12, 90th percentile approximately 0.181, 95th percentile approximately 0.48, and 99th percentile approximately 2.3. Because hazards are not probabilities, values above 1 are admissible. The deployment logic proposed in the synthesis is percentile-based triage for post-discharge interventions such as enhanced education, early follow-up, medication reconciliation, and telehealth reminders.

The limitations are explicit. The data are restricted to four Michigan VAMCs and to readmissions linked to non-compliance. High-cardinality clinical variables such as ICD-9, principal diagnosis, and treating specialty were excluded from the final model because of degrees-of-freedom constraints. External validation and conventional discrimination or calibration summaries such as AUC or Brier score are not reported. This suggests that, within veteran healthcare, VeteranAD can also function as a broad shorthand for VA-oriented predictive workflows rather than only for Alzheimer’s disease screening.

Overall, VeteranAD is best understood as a domain-sensitive label. In autonomous driving it refers to a specific perception-in-plan architecture with anchor-conditioned attention and autoregressive planning (Zhang et al., 15 Aug 2025). In veteran healthcare analytics it has been used for longitudinal Alzheimer’s screening and for operational risk stratification pipelines grounded in EHR notes, eye tracking, or frailty-based survival analysis (Li et al., 2023, Sriram et al., 2023, Ajorlou et al., 2014). For precise scholarly communication, the term therefore benefits from an explicit qualifier such as VeteranAD (autonomous driving) or VeteranAD (veteran health analytics).