We conducted a single-center, prospective pilot cohort study to evaluate the feasibility and usability of a smartphone or web-based self-monitoring system for patients after brain tumor surgery during outpatient follow-up.

The study protocol was approved by the institutional review board of Korea University Guro Hospital (2025GR0177). Written informed consent was obtained from all participants prior to study enrollment. To ensure privacy and confidentiality, all collected data were rigorously deidentified, and patient-level information was stored on a secure, encrypted clinical research server accessible only to authorized personnel. Participants did not receive any financial compensation for their involvement in this study.

Eligible participants were adults discharged after brain tumor surgery at our institution between April 2025 and October 2025. The participant eligibility criteria are presented in Textbox 1.

Of 100 patients approached, 64 (64%) enrolled; after excluding those with <3 total submissions due to access issues (eg, older devices and unstable connections) and 2 midstudy withdrawals, 30 (30%) participants completed the study. Baseline characteristics (enrollment, attrition, sex, and diagnosis) are summarized in Figure 1 and Table 1.

“Clinical deterioration” during follow-up was a composite clinical end point, met when any of the following objective criteria were observed: (1) a decline of ≥20 points in the Karnofsky Performance Status Scale [20,21], (2) new or progressive neurological deficits verified on clinical examination, (3) radiographic progression of intracranial lesions on serial neuroimaging, or (4) clinically significant abnormalities in hematologic or biochemical laboratory results indicating disease progression or systemic decline. Events were ascertained only when patients visited the outpatient clinic or were admitted to the emergency room and were verified by the treating neurosurgeon. To prevent information bias, the treating neurosurgeon adjudicating the clinical deterioration was strictly blinded to the patients’ app-submitted symptom scores. The adjudication relied solely on objective electronic medical records, neuroimaging reports, and standardized clinical examinations obtained during these hospital visits. The “deterioration day” was strictly defined as the date of the patient visit (outpatient or emergency room) during which the objective end point was confirmed and recorded by the physician, rather than a retrospectively estimated date of symptom onset based on patient recall. The remaining participants were classified as having no deterioration over the observation period.

Participants accessed a secure, credentialed web application providing three core functions: (1) Daily symptom check—7 sequential pages of brain-disease–related yes or no items (4-10 items per page), with an automatically computed same-day score based on the number of checked items; (2) “My score”—a time-series visualization of daily scores to support self-tracking; and (3) Free-text inquiries—patient-entered concerns routed to the attending neurosurgeon for in-app responses. The overall user interface and task flow are shown in Figure 2.

At discharge, a research coordinator provided onboarding (log-in, tutorial, and practice submission). Ongoing user support was provided through KakaoTalk (Kakao Corp), a widely used mobile instant-messaging app in South Korea [22,23]. Participants were encouraged to complete at least 1 survey per week beginning the day after discharge. One SMS text message reminder was sent weekly (Monday or Tuesday). Outpatient visits proceeded per standard care; no clinical management changes were mandated beyond app use.

Each daily submission comprised 51 binary items (checked=1 and unchecked=0). The total score on day t was defined as follows: We also computed 7 subcategory subscores as within-category sums. A clinically curated set of “core” items, C = {6,8,12,13,14,22,23,26,29,30,32,33,34,49}, was prespecified as higher-risk symptoms. The complete list of the 51 symptom items, their category assignments, and whether each item was flagged as “higher risk” is provided in Table S1 in Multimedia Appendix 1.

The higher-risk symptoms (eg, confusion and seizures) were designated a priori as red-flag signs. These symptoms were defined as highly objective indicators of acute neurological deterioration that may pose an immediate threat to patient safety if not evaluated promptly. Although the rule-based alert algorithm is still under development, these red-flag signs were predefined as conditions requiring immediate clinical attention, even in the absence of other symptomatic changes. Clinically, these red-flag signs are operationalized to completely bypass the standard cumulative score rules (R1-R4). If a patient endorses any red-flag item, the system is designed to trigger an immediate on-screen warning instructing the patient to seek urgent medical care (eg, visiting the emergency room), independent of the rule-based thresholds. This red-flag designation is orthogonal to the core set C and was used to guide urgent clinical triage.

To reduce within-patient correlation and address class imbalance, the unit of analysis was each patient’s last 3 consecutive submissions (“last-window”: with most recent). For patients with adjudicated clinical deterioration, submissions up to “deterioration day” +1 were retained, and the last window was labeled positive (deterioration=1); otherwise, the window was labeled negative (deterioration=0). Patients with <3 submissions were excluded from rule analysis.

We evaluated 4 simple, interpretable rule families as candidate digital alerts:

Thresholds were tuned on observed, discrete values in the last-window dataset using grid search: (1) R1: T₁ over empirical integer of x_i2, (2) R2: T₂ over empirical integer of , (3) R3: (c_min, n_min)∈{(1,0),(1,1),(1,2),(2,0)}, and (4) R4: K∈{2,3,4}.

The primary selection criterion was area under the receiver operating characteristic (AUROC); ties were broken by F₁-score and then by the positive predictive value (PPV). Final tuned parameters for each rule are reported in Table 2, with performance summarized in Figure 3.

For each rule, we computed sensitivity, specificity, PPV, negative predictive value, F₁-score, accuracy, AUROC, and area under the precision-recall curve using the binary rule outputs on the last-window sample. Uncertainty was quantified with patient-level bootstrap resampling (5000 replicates) and percentile 95% CIs for each metric. Because our analytic approach restricted the dataset to a single “last-window” observation per patient (ie, analyzing only the final interval before an event or the end of follow-up), each participant contributed exactly 1 independent data point to the evaluation dataset. Consequently, within-patient clustering was inherently avoided in this analysis. Analyses focused on discrimination and precision-recall trade-offs appropriate for rare adverse events in small pilot cohorts. The Python code used for these analyses is available in Multimedia Appendix 2.

To address the missing data resulting from study attrition, we used a complete-case analysis approach for evaluating the alert rules. Because the primary objective of this pilot study was to establish initial technical feasibility and conduct rule-based parameter tuning rather than population-level causal inference, imputation of missing longitudinal submissions was not performed. To assess potential selection bias, baseline demographic and clinical characteristics were compared between the included participants (≥3 submissions) and excluded participants (<3 submissions) using the Mann-Whitney U test for continuous variables and chi-square or Fisher exact tests for categorical variables as appropriate. Because the entire last-window cohort was used for both threshold optimization (grid search) and performance evaluation without an independent holdout test set, the reported metrics represent resubstitution (apparent) performance.

We administered a brief questionnaire, including the System Usability Scale (SUS) and additional items on use frequency and initiation triggers. The survey link was distributed via KakaoTalk; responses were anonymized. A total of 27 responses were received; after excluding 4 respondents who reported never using the app, 23 responses were analyzed. SUS scoring followed standard procedures and is reported alongside satisfaction ratings.

Among patients with ≥3 total submissions (analysis cohort; N=30), the mean weekly submission rate was 1.70 surveys per week (range 0.32-6.90); individuals contributed a median of 8.5 (IQR 3-143) submissions over the observation period. The interval between the last 2 submissions (i1→i2) had a median of 17.5 (IQR 15.75; mean 18.4; range 1-105) days, indicating substantial variability and a tendency toward longer gaps near the end of follow-up for some participants. The most recent total score averaged 2.43 points (median 1; range 0-11), and the patient-level mean total score across all submissions averaged 2.21 points (median 1.11; range 0-12.5). Descriptive statistics for use and scores are summarized in Table 3.

A comparison of baseline characteristics between the 30 included participants (completers) and the 34 excluded participants (noncompleters) is presented in Table S2 in Multimedia Appendix 1. There were no statistically significant differences between the 2 groups in terms of age (P=.08), sex (P=.19), or diagnosis categories (P=.36). This lack of significant difference suggests that the initial study attrition was primarily driven by external technological or access issues rather than distinct clinical severity or tumor phenotypes. However, the excluded cohort was significantly older than the included cohort, with a median age of 63.0 years compared to 57.0 years (P=.04). There were no significant differences between the two groups in terms of sex (P=.19) or diagnosis categories (P=.36).

We conducted grid searches to tune thresholds for the 4 prespecified rule families (R1-R4) defined in the Methods section. The optimized parameters are reported in Table 2. For visualization, AUROC was plotted against the candidate thresholds for R1 (current total score, T₁) and R2 (net increase, T₂); both curves showed clear internal maxima, with peak performance at T₁=5 and T₁=2, respectively (Figure 3A). These findings support the face-valid expectation that (1) a moderately elevated current score and (2) short-horizon net worsening each contribute usefully to discrimination.

At the optimized settings, R2 (net increase) showed the best overall balance (AUROC 0.875; sensitivity 0.833; specificity 0.917; PPV 0.714; negative predictive value 0.957; F₁-score 0.769; accuracy 0.900), outperforming other rules on discrimination and F₁-score while preserving high specificity (Figure 3B; Table 4). R1 performed second best (AUROC 0.792; F₁-score 0.667; accuracy 0.867), while R3 and R4 traded sensitivity for specificity (R3: specificity 1.000; sensitivity 0.333; AUROC 0.667 and R4: specificity 0.958; sensitivity 0.333; AUROC 0.646), indicating more conservative alerting with reduced case finding. Bootstrap 95% CIs for sensitivity, F₁-score, and AUROC are shown as error bars in Figure 3B. It is important to note that due to the small sample size and the low number of clinical deterioration events (n=6), these CIs are wide. This indicates that the point estimates (eg, the AUROC of 0.88 for R2) are statistically unstable and should be interpreted with caution. Nevertheless, the point estimates for R2 remain consistently highest across metrics, supporting the value of short-horizon worsening (change from the immediately prior submission) as a digital signal of deterioration.

Item-wise yes-response rates were contrasted between patients with adjudicated deterioration and those without (Figure 4). When grouped by the 7 symptom categories, the “cognition and sleep” and “other symptoms” domains showed notably higher affirmative rates in the deterioration group. At the item level, “memory impairment,” “decreased general condition,” and “muscle weakness” ranked among the largest between-group differences. Conversely, several items predesignated as “core” due to presumed clinical severity were infrequently endorsed in both groups, limiting their utility as direct triggers in this dataset. Taken together, these patterns indicate that comparatively softer but commonly reported symptoms may carry earlier signal for deterioration than rarer, high-severity items.

Of 27 questionnaire respondents, 23 (85%) reported at least 1 instance of app use and were included in the SUS analysis. The SUS total score distribution is shown in Figure 5A (mean 84.0, median 90.0), indicating generally favorable perceived usability. Participants with only 1 submission exhibited a lower SUS distribution than those with ≥2 submissions, consistent with lower adherence among users reporting poorer usability. To probe domain-level patterns, item means were compared between the 1-submission and ≥2-submission groups (Figure 5B): differences were modest for standard (odd-numbered) items, while reverse-scored (even-numbered) items were more negative in the single-submission group. This indicates that infrequent users tend to perceive higher complexity or lower consistency in relation to app use.

Multiresponse questions on (1) reasons for continued use and (2) triggers to initiate a submission are summarized in Figures 5C and 5D. Participants with higher submission counts more frequently selected “personal will or habit” and “perceived recovery benefit” as reasons for sustained use. For initiation triggers, higher-adherence users more often endorsed cues embedded in daily routines. Together, these findings imply that emphasizing perceived clinical usefulness and integrating reminders into routine time points (eg, after lunch) may enhance adherence in future iterations of the intervention.

This preliminary study demonstrated the feasibility and patient acceptability of a rule-based smartphone and web application for self-monitoring in individuals recovering from brain tumor surgery. Following major neurosurgical procedures, even minor physical or neurological changes can provoke significant anxiety [24]. By providing a structured, medically validated alternative to unverified online information, this platform supports objective symptom tracking and bridges the gap between intermittent in-person clinic visits [25-29]. The high usability scores (mean SUS 84.0) suggest that patients are willing to engage with digital health tools during the vulnerable postoperative period, a willingness likely bolstered by the personalized oversight of their treating neurosurgeon. Furthermore, our evaluation of simple heuristic rules, particularly the short-horizon worsening rule (R2), demonstrated preliminary discriminative potential for capturing early signals of clinical deterioration.

While the technical performance of these heuristic rules demonstrates initial feasibility, excellent technical accuracy alone is insufficient for a viable clinical decision support system. A critical challenge for real-world implementation is the risk of alert fatigue. Generating an excessive number of false alarms can easily overwhelm clinical staff and disrupt workflows. Therefore, future deployments of this system will use a structured, tiered triage protocol. Rather than directly interrupting the attending neurosurgeon, a breached cumulative threshold could trigger an automated recommendation for the patient to reschedule a clinic visit or route the alert to a specialized triage nurse for preliminary review. Conversely, prespecified “red-flag” symptoms (eg, seizures) are operationalized as distinct medical emergencies; endorsing these items bypasses the cumulative score rules entirely to provide instantaneous safety instructions, ensuring critically urgent cases receive immediate attention.

From a technical standpoint, the current platform evaluates simple, prespecified threshold rules, which serve merely as a foundational step. The long-term vision of this project is to establish an advanced artificial intelligence–driven early warning framework [8,30,31]. By aggregating longitudinal patient-reported symptom trajectories and eventually integrating continuous physiological data from wearable devices [32-35], future machine learning algorithms could autonomously identify subtle, multivariate deviations. This transition represents a paradigm shift from a passive, provider-centered framework to an active, patient-engaged model of postoperative neurosurgical care [36-40].

As with all pilot studies, several limitations should be acknowledged. First, the small sample size (N=30) and the low number of clinical deterioration events (n=6) mean that the discriminative performance metrics—including the high AUROC observed for R2—are statistically unstable and exhibit wide CIs. These results must be interpreted strictly as preliminary feasibility signals rather than definitive validations. Additionally, to accommodate data sparsity during this pilot phase, we used a simplified “last-window” heuristic and a complete-case analysis. This approach does not fully leverage the longitudinal nature of the data or account for within-patient correlation. Future investigations must address these analytical limitations by using advanced longitudinal methodologies, such as mixed-effects models or time-series machine learning algorithms, in larger independent cohorts. Furthermore, defining the clinical event by the actual hospital visit date introduces potential lead-time bias. Because patients may delay seeking care despite experiencing worsening symptoms, the observed interval between an app threshold breach and the clinical event may overestimate the system’s true early warning capability. Additionally, our threshold parameters were optimized and evaluated on the exact same dataset. Consequently, our findings represent apparent (or resubstitution) performance that is inherently subject to optimistic bias. These rule thresholds must be rigorously evaluated and calibrated on independent, external test sets in future larger-scale studies.

Second, the substantial attrition rate (34/64, 53%) introduces potential selection bias. Although baseline clinical characteristics did not differ significantly between completers and noncompleters, the dropout group tended to be older (median 63.0 vs 57.0 years; P=.04). The platform currently targets patients with mild symptoms capable of independent daily activities, and the high usability scores likely reflect a technologically adept subgroup. However, patients with brain tumors frequently experience postoperative cognitive deficits, visual impairments, or motor limitations that hinder standard smartphone interaction. To address these significant digital literacy and physical barriers, future iterations will incorporate tailored adaptations, including caregiver proxy-reporting functionalities and natural conversational (chatbot) interfaces currently under development. Advanced statistical methods such as multiple imputation should also be used to handle missing data robustly.

Finally, this study was conducted at a single tertiary medical center in South Korea, an environment with exceptionally high smartphone penetration, reliable digital infrastructure, and specific health care follow-up protocols. Consequently, the high engagement rates and feasibility observed may not directly translate to other global health care systems or populations with different socioeconomic constraints. Multicenter validations across diverse geographic contexts are required to externally validate these digital self-monitoring tools.

Our preliminary report provides early evidence that smartphone-based self-monitoring systems are both feasible and acceptable among patients with brain tumors. Beyond usability, this research lays the groundwork for the next phase of development—building predictive, artificial intelligence–powered early warning algorithms that can autonomously identify high-risk symptom patterns and deliver timely alerts. Such an approach has the potential to redefine postoperative neurosurgical care, enhance digital health care accessibility, and ultimately improve patient safety and long-term outcomes.