Van Gogh Analysis Reveals Visual Complexity Decoupled From Written Sentiment Before Death

A computational investigation combining mathematical image processing and natural language analysis has identified a quantifiable behavioral divergence in the final creative output of Vincent van Gogh, offering a retrospective model for tracking severe psychiatric crisis. Researchers from Fukushima Medical University School of Medicine and the DMC Institute for Life Science analyzed 388 cataloged paintings alongside personal correspondence produced across the artist’s last five clinical phases between 1888 and 1890. The findings demonstrate that while geometric brushstroke complexity rose to sustained peaks during periods of acute psychological instability and remained elevated until his death, the negative emotional valence expressed in his written correspondence dropped sharply during his final weeks in Auvers-sur-Oise. This statistical decoupling between visual execution and verbal expression provides an objective behavioral pattern that parallels the clinical phenomenon of terminal resolve, where individuals approaching suicide exhibit apparent emotional stabilization despite profound underlying distress.

The study addresses a longstanding limitation in computational psychiatry and digital phenotyping, which seek to extract objective behavioral signatures from continuous streams of personal data such as daily smartphone interactions and social media text. Contemporary psychiatric assessment continues to rely heavily on subjective self-reporting and episodic clinical interviews, modalities that can fail to capture latent suicidal intent or rapidly fluctuating affective states. While digital phenotyping has advanced by monitoring linguistic markers, unimodal tracking remains vulnerable to deceptive symptomatic remission, wherein an individual’s outward verbalizations normalize while severe distress persists. Developing and validating multimodal frameworks capable of concurrently evaluating independent behavioral channels requires dense, chronologically cataloged data gathered over prolonged periods of crisis. Because collecting continuous multimodal records from living patients over years presents severe practical and ethical hurdles, the researchers turned to historical pathography as an empirical paradigm.

Vincent van Gogh’s prolific late period provided an exceptionally dense single-subject corpus to test whether independent behavioral channels synchronize during distress and diverge prior to suicide. This synchronized output generated an uninterrupted record reflecting both external psychomotor execution and internal emotional reflection across five clinically distinct phases. By segmenting this 29-month timeframe according to objective biographical milestones—including geographical moves between Arles, the Saint-Paul-de-Mausole asylum in Saint-Rémy, and Auvers-sur-Oise—the research team established an empirical baseline to measure how structural patterns in visual art co-varied with the emotional tone of written correspondence across consecutive psychiatric transitions.

To evaluate the physical structure of the artwork without interference from color choices, the investigators developed a standardized computational pipeline applying fractal dimension analysis to high-resolution digital reproductions. Each image was standardized to a uniform color space and resolution before being converted to grayscale, after which a Canny Edge Detection algorithm isolated the structural contours of the artist’s impasto brushstrokes and compositional lines. The team then quantified the geometric complexity and self-similarity of these edges using the box-counting method, a mathematical algorithm implemented via Python, OpenCV, and NumPy. Square grids with decreasing side lengths ranging from 21 to 210 pixels were overlaid onto each binary edge image to count edge-containing boxes. The fractal dimension was extracted from the absolute slope of the least-squares linear regression line on a log-log plot, with all 388 paintings exceeding an R-squared goodness-of-fit value of 0.95, confirming that the late brushwork adhered to genuine fractal-like scaling properties rather than random visual noise.

The longitudinal framework partitioned the corpus into five chronological periods defined a priori by biographical events rather than computational patterns. Period I, the Early Arles phase running from February to October 1888, captured van Gogh’s arrival in southern France, marked by chromatic exploration and the incorporation of Japanese artistic principles. Period II, the Yellow House phase spanning November and December 1888, encompassed rising interpersonal agitation during his cohabitation with Paul Gauguin, culminating in his initial breakdown and ear mutilation. Period III, lasting from January to April 1889, represented post-crisis recuperation and fluctuating stability during repeated hospitalizations in Arles. Period IV covered his voluntary confinement at the Saint-Paul-de-Mausole asylum in Saint-Rémy from May 1889 to May 1890, a phase of protracted isolation. Finally, Period V spanned June and July 1890 in Auvers-sur-Oise, representing his concluding weeks of frantic artistic production under acute existential vulnerability.

In parallel with the visual modeling, the researchers quantified emotional tone across the correspondence using the Valence Aware Dictionary and sEntiment Reasoner, or VADER, an established natural language processing tool designed for creative and informal texts. Rather than relying solely on the broad compound metric, the researchers isolated the specific proportions of positive and negative sentiment scores to distinguish whether affective shifts stemmed from active surges in negative feelings or passive reductions in positive expression. These granular linguistic metrics were then paired chronologically with the corresponding period’s fractal dimensions to compare internal sentiment directly against external compositional execution.

Statistical evaluation using one-way analysis of variance demonstrated that the geometric complexity of van Gogh’s paintings varied significantly across the five clinical periods, yielding an F-ratio of 15.57 with a p-value below 0.001 and an effect size of 0.14. As shown in the study’s trajectory, the mean fractal dimension dropped to its lowest recorded point during Period II, reaching a value of 1.4473 in representative works such as the Portrait of Armand Roulin. Following this nadir, the structural complexity rebounded during the hospital recuperation of Period III to baseline levels, before escalating significantly during Periods IV and V. Tukey’s post hoc comparisons confirmed that the Yellow House days exhibited a significant reduction in compositional complexity compared to every other phase, whereas the asylum confinement and final Auvers period maintained sustained elevations in structural density and intricate brushstroke execution.

The computational analysis of the letters revealed a contrasting dynamic across the same timeline, highlighted by marked shifts in negative emotional expressions alongside striking stability in positive language. The positive sentiment score showed no statistically meaningful variation across the five periods, yielding an F-ratio of 0.91 and a p-value of 0.459, indicating that van Gogh maintained a consistent baseline of positive expression in his writing regardless of psychological circumstances. Conversely, negative sentiment exhibited pronounced fluctuations, producing a significant F-ratio of 6.21 with a p-value below 0.001 and an effect size of 0.09. Negative sentiment fell during Period II, aligning with the initial optimism of his collaboration with Gauguin, but surged sharply during Periods III and IV during his hospitalizations and asylum stay. Crucially, this elevated negative sentiment dropped markedly in Period V, demonstrating that his written negativity receded in the final weeks preceding his death.

To evaluate whether visual structure and linguistic expression operated synchronously, the investigators computed Pearson correlation coefficients on the aggregated phase means across the periods. When analyzing the first four phases, from the Early Arles period through the Saint-Rémy asylum stay, the correlation coefficient between mean fractal dimension and mean negative sentiment reached 0.81. While this strong association did not achieve formal statistical significance due to the small sample size of four aggregated historical intervals, the large effect size indicated that psychomotor agitation and verbalized negative affect tracked each other closely during the initial phases of distress. However, when the final Auvers-sur-Oise phase was included in the calculation, the correlation coefficient dropped sharply to 0.47. This mathematical collapse captures the multimodal decoupling phenomenon, illustrating that while his brushstroke complexity remained near peak levels, his written language paradoxically normalized.

The authors contextualized these computational measurements within neuropsychological principles and modern complexity theory. The peak in fractal dimension observed during the late asylum period corresponds with previous physics literature demonstrating that van Gogh’s late masterpieces exhibit luminance fluctuations matching the mathematical structure of Kolmogorov fluid turbulence. Neuropsychologically, the preservation of elevated structural complexity indicates that even while experiencing profound psychological instability, the artist retained high cognitive organization, fine motor control, and creative focus. Consequently, the turbulent structural execution in his late canvases represents an objective behavioral correlate of internal emotional intensity. Interpreting these patterns through dynamic systems theory, the authors propose that psychiatric trajectories represent multidimensional complex systems where non-linear shifts between independent behavioral outputs reflect severe systemic strain rather than simple linear degradation.

The study highlights translational potential for contemporary clinical suicide prevention while acknowledging substantial epistemological caveats. Modern digital health platforms could capture analogous multimodal data by integrating text from messaging applications with fine motor kinetics recorded through smartphone touchscreens, wearable actigraphy sensors, or digital art therapy tools. If automated monitoring algorithms detect a decoupling pattern—where explicit linguistic sentiment normalizes while underlying psychomotor entropy remains critically elevated—the system could trigger clinical alerts for rapid intervention before an individual acts on latent suicidal intent. However, the authors emphasize that diagnostic classifications for van Gogh remain contested, with historical hypotheses spanning bipolar disorder, temporal lobe epilepsy, borderline personality traits, and substance-induced psychosis. Because the digital markers cannot be tied to a verified diagnostic category, they should be viewed as non-specific behavioral correlates of severe psychological turmoil.

The researchers also detailed critical methodological limitations inherent to retrospective single-subject investigations lacking independent clinical ground truth. The absence of standardized psychiatric diagnostic evaluations or matched control trajectories—such as historical cohorts experiencing acute distress that did not culminate in suicide—prevents confirming whether multimodal decoupling is specific to terminal resolve or occurs across other psychiatric fluctuations. Furthermore, the sentiment analysis relied on English translations rather than original Dutch and French texts, introducing potential translation biases, cultural shifts, and the loss of subtle nuances that rule-based tools like VADER cannot decipher. Archival dating uncertainties for specific artworks and letters also introduce temporal noise into period alignments. While future studies must validate multimodal frameworks in prospective clinical cohorts using context-aware transformer language models, this historical reconstruction establishes an empirical foundation for tracking non-linear behavioral signatures across multimodal data streams.

Subject of Research: Multimodal digital phenotyping of psychological distress using historical artistic complexity and linguistic sentiment

Article Title: Reconstructing digital phenotyping of psychological distress: integrating visual complexity and linguistic sentiment

Article References: Yokooka, R., Suzuki, M., Sekine, H., Takenoshita, S., Endo, S., Maejima, Y., & Shimomura, K. (2026). Reconstructing digital phenotyping of psychological distress: integrating visual complexity and linguistic sentiment. Frontiers in Psychiatry, 17, Article 1946175. https://doi.org/10.3389/fpsyt.2026.1946175

Image Credits: AI Generated

DOI: 10.3389/fpsyt.2026.1946175

Keywords: digital phenotyping, computational psychiatry, Vincent van Gogh, fractal dimension, VADER sentiment analysis, visual complexity, box-counting method, Canny edge detection, psychomotor turbulence, terminal resolve, multimodal decoupling, historical pathography, natural language processing, Saint-Rémy asylum, Auvers-sur-Oise, affective dynamics