Acute mountain sickness is far more than an inconvenience. Headache, dizziness, nausea, vomiting, and fatigue can degrade judgment and performance, and in severe cases the condition can progress to high-altitude cerebral edema or high-altitude pulmonary edema, both of which are life threatening. As modern transportation makes it possible to fly from Beijing to the Tibetan plateau in under five hours, the number of people exposed to sudden, severe hypoxia has risen sharply. Yet the tools available to predict who will fall ill remain frustratingly limited, often requiring hypoxic gas generators, hypobaric chambers, blood biomarkers, or expensive imaging equipment that is unrealistic for field deployment.

The research team, drawn from Tsinghua University and collaborating institutions, recruited 48 healthy low-altitude residents aged 18 to 31 years with no high-altitude exposure in the preceding year. After strict screening for cardiovascular disease, chronic respiratory conditions, hypertension, and medications affecting cardiopulmonary function, 44 participants completed the full protocol. At 50 meters above sea level, each volunteer performed a standardized step test involving three minutes of seated rest, five minutes of stepping at 22.5 steps per minute on a platform 35 centimeters high for women and 40 centimeters for men, and four minutes of seated recovery. Throughout the protocol, a finger-clip pulse oximeter continuously recorded heart rate and peripheral oxygen saturation, while investigators monitored pulse waveforms to discard any artifactual readings.

Days later, the same participants boarded a commercial flight from Beijing to Lhasa, followed by a bus transfer to an experimental base, with all arrivals synchronized between 14:00 and 15:00 to minimize travel-fatigue confounds. The following morning, 18 hours after arrival, researchers administered the Lake Louise Score, the widely used self-report instrument for AMS, under blinded conditions in an independent space. Nineteen of the 44 participants met the diagnostic criteria for AMS, defined as a score of 3 or greater with headache plus at least one additional symptom.

Back at low altitude, the physiological data told a clear story. Variables captured during exercise and recovery, including estimated maximal oxygen uptake, exercise oxygen saturation, recovery oxygen saturation, exercise heart rate, recovery heart rate, and a composite step index, were all significantly correlated with subsequent AMS severity. In contrast, resting heart rate and resting oxygen saturation showed no meaningful association. This distinction matters physiologically: a resting baseline rarely exposes hidden limitations in ventilatory or cardiovascular compensation, whereas the added oxygen demand of exercise can reveal subtle deficits in how efficiently the body shuttles and utilizes oxygen, mimicking in miniature the stress that sudden altitude exposure imposes.

To translate these observations into a classification framework, the researchers applied k-means clustering, an unsupervised machine learning algorithm that groups individuals based on similarity of features without requiring predefined labels. After standardizing the data and reducing dimensionality with principal component analysis, the team tested multiple combinations of physiological variables. The strongest and most accurate stratification emerged from four indicators: estimated VO2max, exercise oxygen saturation, recovery oxygen saturation, and the step index. This combination yielded a silhouette coefficient of 0.767, indicating a strong two-cluster structure, and achieved a within-cohort accuracy of 93.18 percent when evaluated against the actual Lake Louise classifications.

Perhaps most striking was the sensitivity of the approach. All 19 participants who developed AMS were assigned to the AMS-susceptible cluster, and none of the 22 participants in the non-susceptible cluster developed symptoms. The authors are careful to note that these figures represent within-cohort clustering performance rather than validated predictive accuracy in an independent sample, and that no false-negative assignments in a new cohort cannot be guaranteed. Nevertheless, the effect sizes separating the two clusters were substantial, with Cohen’s d values ranging from 0.99 for estimated VO2max to 1.75 for the step index, and the Lake Louise Score itself differed markedly between clusters.

The team went to considerable lengths to confirm the clustering was not an artifact. Bootstrap resampling across 3000 replicates produced a median Jaccard stability index of 0.803, well within the range considered stable. Alternative algorithms including fuzzy c-means, partitioning around medoids, Gaussian mixture models, spectral clustering, and Ward hierarchical clustering broadly reproduced the same partition. When estimated VO2max was deliberately perturbed with realistic measurement error, or excluded entirely, the core structure persisted with only modest degradation. Sex, which some prior studies have linked to AMS susceptibility, showed no significant association with cluster membership, and statistically removing sex-related differences in estimated fitness left every participant in their original cluster.

The most influential single variable turned out to be the step index, a simple composite derived from recovery heart rates and exercise duration. Though modest alone, it combined powerfully with exercise oxygenation measures to separate the two physiological phenotypes. The authors suggest this reflects the fundamental importance of cardiorespiratory reserve: people whose bodies recover quickly from submaximal exertion and who maintain oxygen saturation under load appear better equipped to handle the abrupt hypoxic burden of rapid ascent. Three participants classified as susceptible but who scored below the AMS threshold may represent individuals with genuinely compromised reserves or, alternatively, conservative symptom self-reporting that underestimated their true Lake Louise Scores.

The implications for public health and occupational medicine are considerable. Mountaineers, military personnel, railway and construction workers, pilgrims, and ordinary tourists all stand to benefit from a screening method that requires nothing more than a step platform, a pulse oximeter, and twelve minutes of time. Unlike hypoxic chamber tests or blood-based omics panels, the protocol is easily standardized and could plausibly be administered at worksites, travel clinics, or recruitment centers. The authors emphasize that the approach remains exploratory and requires validation in larger, more diverse cohorts spanning different ages, health statuses, ascent profiles, and altitude targets, and that severe outcomes such as high-altitude cerebral or pulmonary edema were not represented in this young, healthy sample. Still, the study demonstrates that meaningful physiological structure emerges from a test simple enough to be administered almost anywhere, offering a glimpse of a future in which altitude illness risk can be identified and mitigated before the first symptom ever appears.

Subject of Research: Prediction of acute mountain sickness susceptibility using low-altitude step test data and unsupervised machine learning clustering

Article Title: Clustering analysis of acute mountain sickness susceptibility among young adults during rapid ascent using low‐altitude step test data

Article References: Clustering analysis of acute mountain sickness susceptibility among young adults during rapid ascent using low‐altitude step test data. (n.d.). https://doi.org/10.14814/phy2.71091

Image Credits: AI Generated

DOI: 10.14814/phy2.71091

Keywords: acute mountain sickness, step test, k-means clustering, hypoxia, high altitude, machine learning, oxygen saturation, cardiorespiratory fitness, Lake Louise Score, rapid ascent, pulse oximetry, physiological screening