Machine Learning Evaluates Dementia-Friendly Architectural Sightlines in Early Building Layouts

A computational framework combining machine learning and geometric isovist modeling can objectively evaluate early-stage architectural floor plans against established dementia design principles, according to a study published in Architectural Intelligence. Developed by Feras Alsaggaf, AnneMarie Eijkelenboom, Martijn Lugten, and Michela Turrin, the proof-of-concept system analyzes spatial configurations to predict visual access quality before layouts advance to costly developmental and expert validation phases. By translating qualitative care guidance into quantitative spatial indicators, the model determines whether building floor plans provide sufficient visual connections between key living zones and essential destinations such as kitchens and sanitary facilities. Testing on residential layout configurations demonstrated that the multi-output machine learning model achieved classification accuracies between 84 percent and 87 percent across individual visual access categories, alongside a 72 percent subset accuracy when predicting multiple spatial criteria simultaneously. The findings present a practical data-driven pathway for architects to rapidly benchmark alternative floor layouts during preliminary schematic design, directly addressing the sensory and cognitive navigation challenges experienced by residents living with dementia.

Dementia is a progressive neurodegenerative disease marked by memory loss and a diminished capacity to complete daily tasks independently, standing as the primary global driver of disability and dependency among older demographics. The physical progression of dementia impairs path integration, cognitive mapping, and sensory perception across vision, hearing, and spatial awareness, causing distorted interpretations of environmental stimuli that can severely erode an individual’s personal identity and autonomy. In residential care facilities, the preservation of personal autonomy depends heavily on an occupant’s ability to navigate their physical surroundings safely, exert choice, and maintain pre-dementia habits and values. Layouts conceived without accounting for these specific spatial needs create structural barriers to orientation and movement, triggering emotional distress, anxiety, and behavioral agitation when residents lose their way. Architecture that intentionally supports human navigation—encompassing both cognitive wayfinding decisions and physical locomotion—can significantly alleviate cognitive confusion, allowing individuals to maintain higher levels of functional independence and daily social engagement throughout long-term care environments.

A foundational guideline within specialized design frameworks, such as the Environmental Assessment Tool developed by Fleming, is the principle that enables occupants to see and be seen across communal spaces. In the formal scoring rubric of the Environmental Assessment Tool, visual sightline metrics account for 33 of 112 total points, representing 29.5 percent of the overall evaluation weight and underscoring the vital role visual access plays in dementia-inclusive construction. Direct sightlines reduce the cognitive burden of route planning by compensating for degraded internal spatial memory, enabling individuals to identify destinations immediately without needing to recall complex floor plans. Prior studies indicate that unobstructed views between frequently occupied areas, such as lounges, dining spaces, kitchens, and toilets, foster spontaneous social interactions with peers and caregivers while enhancing residents’ confidence to explore their living spaces. Simultaneously, transparent lines of sight assist nursing staff by streamlining visual monitoring and rapid supervision, allowing care teams to intervene effectively when an older adult experiences spatial disorientation.

Architectural researchers have previously explored computational methods to study spatial perception and sightlines, frequently relying on visibility graph analysis derived from space syntax and the classical isovist technique formulated by Clifford Tandy in 1967 and expanded by Michael Benedikt in 1979. An isovist mathematically defines the entire polygon of space directly visible from a specific observer vantage point, providing an objective descriptor of spatial volume and enclosure. While earlier research employed isovists to evaluate visual connections in specific residential settings or examined visibility graphs to measure accessibility, existing literature exhibited a substantial operational gap regarding standardized measurement schemes for dementia-specific environmental criteria. Prior evaluation methods frequently required evaluators to manually select vantage points, introducing subjective bias and inconsistency because exact furniture placements and focal points evolve across a building’s lifespan. Furthermore, while machine learning surrogate models had been tested on workplace layout optimization, automated computational tools bridging qualitative dementia care guidelines with data-driven early design validation remained largely unexploited.

To overcome these methodological shortcomings, the research team established an automated isovist-based quantification workflow capable of evaluating visual access quality across entire room geometries without relying on subjective observer placement. Setting the living room or central lounge as the primary reference space due to its pivotal role connecting to kitchens, dining rooms, gardens, corridors, and toilets, the protocol deployed a uniform grid of test points spaced one meter apart and set one meter away from boundary walls to approximate a human-scale field of view. From each living room grid vertex, the system radially generated isovists to test for direct geometric line-of-sight intersections with adjacent kitchen and bathroom target boundaries. Layouts were subsequently categorized into visual access quality labels based on the proportion of test points meeting direct sightline conditions: insufficient when visual access was largely blocked, sufficient when partial connections existed, and preferred when most vantage points maintained clear sightlines.

Developing and training the predictive framework required an established spatial database with diverse geometric variations, which the authors derived from the publicly available Swiss Dwellings Dataset version 3.0.0 hosted on Zenodo. The researchers filtered the comprehensive dataset containing tens of thousands of residential units using Python and Pandas, selecting single-story apartments containing distinctly labeled living rooms, corridors, and dedicated kitchens, while restricting the sample to dwellings with more than three bedrooms to exclude small studio flats. Following the removal of duplicate floor plans through manual inspection, the final training pool comprised 256 unique apartment layouts. This subset was imported into the Grasshopper parametric modeling environment, where custom computational scripts systematically generated isovist fields across the living room areas, evaluated line-of-sight intersections toward kitchens and nearest toilets, and recorded quantitative visibility percentages to establish balanced ground-truth visual access quality labels for supervised machine learning training.

Alongside visual access labels, the floor plans contained approximately 360 detailed simulation attributes covering environmental acoustics, solar exposure, outdoor views, area dimensions, network connectivity, and geometric features. Because visual access quality is governed by structural walls and spatial partitions rather than climatic variables, the investigators eliminated environmental factors such as sunlight and external sound, narrowing the candidate pool to 70 distinct spatial and topological features. Exploratory data analysis, including histogram distributions and scatter plots with kernel density estimation contours, filtered out weakly correlated variables. For complex network metrics presented in multiple statistical formats, such as connectivity eigen centrality, only the 20th and 80th percentiles were retained to reduce computational overhead, yielding an 18-feature subset that encapsulated layout compactness, perimeter lengths, entrance distances, betweenness centrality, and closeness centrality across the room networks.

To isolate the most effective predictive variables without excessive computational expenditure, the authors implemented forward and backward sequential feature selection using a Random Forest classification algorithm. Forward sequential feature selection identified an optimal seven-feature input subset comprising layout compactness, door perimeter, total number of doors, layout maximum rectangle width, the 20th percentile of bathroom connectivity distance, and both the 20th and 80th percentiles of kitchen connectivity distance. The Random Forest classifier was trained on a 70 percent data split comprising 179 dwellings, with the remaining 30 percent, or 77 dwellings, reserved for independent testing. Hyperparameters were tuned using an exhaustive grid search, and fivefold cross-validation was strictly maintained across the training data to mitigate overfitting risks stemming from the bounded sample size of the architectural dataset.

Performance evaluation was conducted across both individual class metrics and multi-output classification standards, using confusion matrices, precision, recall, F1 scores, subset accuracy, and Hamming loss to measure how effectively the model concurrently predicted separate spatial relationships. In model predictions for the living room to nearest toilet, evaluation reports showed balanced performance with an accuracy of 0.84 and F1 scores between 0.80 and 0.85 across both categories.

Despite these encouraging validation results, the authors highlighted critical methodological boundaries and identified necessary avenues for future research before such tools can be deployed broadly across professional architectural practices. A principal constraint involves the training data source, as the Swiss Dwellings Dataset reflects standard European residential apartment layouts rather than specialized institutional dementia care facilities, meaning findings cannot be directly generalized to differing typologies without retrained or synthetic domain-specific datasets. Additionally, spatial wayfinding encompasses non-visual environmental stimuli, including acoustic resonance, material contrast, and corridor articulation, which require distinct quantification metrics. The authors concluded that realizing the framework’s full potential will necessitate cross-disciplinary collaboration among architects, care professionals, and computer scientists, alongside direct integration into Building Information Modeling platforms to deliver real-time, evidence-based visual guidance during early design exploration.

Subject of Research: Machine learning evaluation of visual access metrics in early-stage dementia-friendly architectural design

Article Title: A machine learning framework for early-stage dementia-friendly architectural design evaluation using visual access metrics

Article References: Alsaggaf, F., Eijkelenboom, A., Lugten, M., & Turrin, M. (2025). A machine learning framework for early-stage dementia-friendly architectural design evaluation using visual access metrics. Architectural Intelligence, 4(1), Article 11. https://doi.org/10.1007/s44223-025-00087-2

Image Credits: AI Generated

DOI: 10.1007/s44223-025-00087-2

Keywords: Dementia-Friendly Architecture, Machine Learning, Visual Access Metrics, Isovist Analysis, Floor Plan Layouts, Wayfinding Support, Environmental Assessment Tool, Random Forest Classifier, Swiss Dwellings Dataset, Topological Features, Multi-Output Classification, Residential Care Facilities, Personal Autonomy, Spatial Configuration, Visibility Graph Analysis, Early Design Validation, Sequential Feature Selection