Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Lily Ingsrisawang has a total of 51 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Learning Risk Factors from Homogeneous and Heterogeneous Survival Data: A Hybrid Machine Learning Framework with Applications to Heart Failure and Leprosy

    by Patcharaporn Porndumnernsawat, Till Daniel Frank, Suree Khemthong, Pramote Luenam, Lily Ingsrisawang

    Background/Objectives: Machine learning-based survival analysis methods often outperform classical approaches but entail greater methodological complexity and computational cost. This study proposes a unified three-component framework for selecting optimal survival modeling strategies by jointly considering data characteristics and research objectives. Methods: The framework comprises variable selection, classical and tree-based machine learning survival models, and performance evaluation using both time-aggregated and time-resolved metrics. It is validated on two distinct datasets of heart failure and leprosy, representing homogeneous and heterogeneous survival structures. Results: It is found that the optimal machine learning approach and variable selection are data dependent. For the heart failure dataset, a bagging survival tree achieves optimal performance, using only six predictors. For the leprosy dataset, the full predictor model yields the best performance, while the computationally efficient random survival forest performs comparably to the more complex bagging approach. Beyond model selection, the framework explicitly incorporates secondary objectives, including parsimony and computational efficiency, and supports the derivation of interpretable decision pathways for risk stratification. Conclusions: Heart failure patients with an ejection fraction below 20% should be considered as the highest risk group, whereas leprosy patients with symptom duration under one month can be regarded as the lowest risk category. In short-term prediction settings (less than 50 days/one year for heart failure/leprosy), the classical Cox proportional hazards model performs comparably to machine learning methods and may be chosen as the preferred method for simplicity. Overall, the proposed multi-component framework enables context-aware model selection by balancing predictive performance, model complexity, and temporal considerations.

  • Open Access

    Overcoming Dengue Underreporting: A Machine Learning Framework for Automated Dengue Notification Based on Electronic Medical Records

    by Warodom Sornsurin, Worasait Suwannik, Panithee Thammawijaya, Till D. Frank, Lily Ingsrisawang

    Hospital notification of dengue cases is frequently delayed, contributing to underreporting. At Thai provincial hospitals the bottleneck is administrative, not diagnostic: delayed discharge coding and clerical steps mean notifications are submitted late or omitted. Existing machine learning (ML) studies require clinician-annotated features absent in passive electronic medical records (EMRs). We propose a dual-pathway EMR framework for ML-based pre-population of surveillance reports. We analyzed 1031 dengue-suspected encounters (508 notifiable) from Surin Hospital, Thailand (2019–2022). Pathway A combines EMR data with manually documented clinical features; Pathway B uses only passive EMR data. Using an 80/20 stratified split, we tested 7 ML methods across 11 feature sets (77 combinations) with SHAP-guided feature elimination and 10-fold cross-validation. On the primary split, Pathway A (CatBoost) achieved an F1 score of 96.62%, 95.10% sensitivity, and 98.10% specificity (10-seed mean F1 95.7% ± 1.4%). Pathway B (LightGBM) achieved an F1 score of 84.42%, 94.12% sensitivity, 75.24% specificity, and a projected negative predictive value of 99.1% at 10% prevalence; 10-seed performance was lower (F1 78.5% ± 4.3%; sensitivity 89.1%). An exploratory LSTM-16 stacking model (F1 83.99%; precision 79.49%) performed best in precision and specificity among Pathway B. Test F1 SD across 10 seeds was ≤5.0% in 75 of 77 combinations. Within this cohort, passive EMR modeling achieved 91–94% sensitivity on the primary split (85–89% across seeds) without additional data collection, supporting the feasibility of ML-based notification in resource-constrained hospitals. Whether this reduces dengue underreporting must be established in prospective, multi-center studies.

  • Open Access

    On a Bayesian multivariate survival tree approach based on three frailty models

    by Patcharaporn Porndumnernsawat, Till D. Frank, Lily Ingsrisawang

    This study compared the performance of the Bayesian multivariate survival tree approach constructed from extended Cox proportional hazard with gamma frailty term, and two shared gamma frailty models with exponential and Weibull baseline hazard function, respectively. A simulation study was applied to evaluate the impact of the baseline hazard function, number of clusters (200, 500, 1000), cluster size (5, 10, 20), and right censoring rate (10%, 50%, 80%) on the performance of classification. We generated 90 clustered survival datasets having correlated failure times and 50 covariates at cluster level and at individual level. Each dataset was resampling 1000 times by selecting clusters at random 70% as training datasets and the rest 30% as the test datasets. The performance of a Bayesian multivariate survival tree approach based on shared gamma frailty models with Weibull distribution provided the highest accuracy. All three models, the accuracy tended to increase with an increase in the cluster size and the number of clusters. The accuracy decreased monotonically with increasing the percentage of censoring rate. In conclusion, the use of the Bayesian multivariate survival tree approach constructed from the shared gamma frailty with baseline hazard function as Weibull distribution was recommended.


  • Importance of Frailty Distributions for Bayesian Multivariate Exponential Survival Trees to Identify Unknown Risk Groups

    by Patcharaporn Porndumnernsawat, Till D. Frank, Lily Ingsrisawang

    The exponential function is commonly used in modeling Cox proportional hazard model. When the lifetime data are correlated and exhibit a clustered structure, the more appropriate statistical methods are studied by combining the methods of multivariate shared frailty model, survival tree, and Bayesian approach. This is demonstrated with an extensive simulation study for time to tooth loss data coming from elderly patients with the aim to identify unknown risk groups. The distribution of frailty term may affect the modeling process in generating Bayesian survival tree. Then, two frailty distributions of gamma and log-normal were focused in multivariate shared exponential survival trees, using Bayesian approach. A total of 45 simulated clustered survival datasets were generated based on number of clusters, cluster size, and right censoring rate. Each dataset contained correlated failure times, and up to 50 covariates at both the cluster and unit levels depending on the selected variables from the stepwise procedure in the Cox proportional hazard model. Next, each dataset was resampled 1000 times, and for each resampled dataset 70\(\%\) of the clusters were used for training and the remaining 30\(\%\) for testing. Across all models, it was found that the classification accuracy increased when the cluster size and number of clusters went up, but consistently decayed when the right-censoring rate increased. Moreover, the Bayesian multivariate survival tree approach incorporating the shared log-normal frailty with an exponential baseline hazard achieved the highest classification accuracy. In conclusion, this model is recommended for its superior classification accuracy across varying conditions. The better performance of the log-normal model highlights the importance of the frailty distribution

  • Predictors of PM${}_{\mathbf{2.5}}$ Concentrations in Bangkok, Thailand: Obtaining Novel Insights for Vertical Measurements by Combining Feature Selection and Deep Learning Approach

    by Jaruwan Mernngurn, Chantha Wongoutong, Till D. Frank, Lily Ingsrisawang

    This study aimed to determine predictors of particulate matter (PM\({}_{2.5}\)) concentrations at different heights (30 m, 75 m, and 110 m) at a monitoring tower in Bangkok, Thailand, a megacity representative of various Asian megacities suffering from air pollution problems. Ground level PM\({}_{2.5}\) air pollution data were used as controls. In a first step, feature selection methods were applied to narrow down the number of relevant predictors. Subsequently, a machine learning algorithm, namely, a deep long short-term memory model, was applied to identify the most relevant predictors. With this approach, we identified the following key PM\({}_{2.5}\) predictors: carbon monoxide for low levels (ground level and 30 m), ozone for medium and high (75 and 110 m) levels, and pressure for high levels (110 m). We also found that using all available features as predictors led to higher prediction errors and, consequently, in general, seems to be a modeling strategy that should not be pursued. Moreover, varying the input data length of the model was a key step to improve prediction accuracy. PM\({}_{2.5}\) signals observed at different vertical levels exhibited height-specific optimal input lengths ranging from 1 to 7 days. Finally, predicting low level (30 m) PM\({}_{2.5}\) concentrations as observed in our vertical PM\({}_{2.5}\) data set and predicting PM\({}_{2.5}\) concentrations from one of the ground level control monitoring sites located nearby a heavy-traffic road could be achieved best with two models that both used carbon monoxide as predictor and a data input of 5 days.


  • Open Access

    Identifying the determinants of tourism receipts of Thailand and relevant determinant-determinant interactions

    by Suree Khemthong, Pramote Luenam, Till D. Frank, Lily Ingsrisawang

    The study examined the determinants that affect tourism receipts in Thailand. To this end, quarterly data from eight main provinces of Thailand from the period 2015–2019 were used and constituted a repeated measures design. Accordingly, a generalized linear mixed model was applied for developing two different random intercept models by treating 1) province, and 2) a combination of province and calendar quarter as cluster-specific effects. It was found that determinants that increased tourism receipts were the number of visitors, the average cost per day, the length of stay of visitors, the presence of low-cost airlines, and a relatively low offence rate. Moreover, an increase in the number of visitors in the fourth quarter produced a higher amount of additional receipts as compared to a similar increase in the first quarter. Specifically, for Thailand attracting high-spending tourists and extending tourist visas for more than 30 days is recommended. Beyond Thailand, uncovering interaction effects as described above may help tourism agencies to focus their limited resources on the determinants that matter.


  • Interplay of xenobiotic-degrading and antibiotic-resistant microorganisms among the microbiome found in the air, handrail, and floor of the subway station

    by Piyanun Harnpicharnchai, Paopit Siriarchawatana, Sermsiri Mayteeworakoon, Lily Ingsrisawang, Somsak Likhitrattanapisal, Lily Eurwilaichitr, Supawadee Ingsriswang


  • Open Access

    Association between out-patient visits and air pollution in Chiang Mai, Thailand: Lessons from a unique situation involving a large data set showing high seasonal levels of air pollution

    by Tunyathron Varapongpisan, Till D. Frank, Lily Ingsrisawang

    Chiang Mai is one of the most known cities of Northern Thailand, representative for various cities in the East and South-East Asian region exhibiting seasonal smog crises. While a few studies have attempted to address smog crises effects on human health in that geographic region, research in this regard is still in its infancy. We exploited a unique situation based on two factors: large pollutant concentration variations due to the Chiang Mai smog crises and a relatively large sample of out-patient visits. About 216,000 out-patient visits in the area of Chiang Mai during the period of 2011 to 2014 for upper (J30-J39) and lower (J44) respiratory tract diseases were evaluated with respect to associations with particulate matter (PM10), ozone (O3), and nitrogen dioxide (NO2) concentrations using single-pollutant and multiple-pollutants Poisson regression models. All three pollutants were found to be associated with visits due to upper respiratory tract diseases (with relative risks RR = 1.023 at cumulative lag 05, 95% CI: 1.021–1.025, per 10 μg/m3 PM10 increase, RR = 1.123 at lag 05, 95% CI: 1.118–1.129, per 10 ppb O3 increase, and RR = 1.110 at lag 05, 95% CI: 1.102–1.119, per 10 ppb NO2 increase). Likewise, all three pollutants were found to be associated with visits due to lower respiratory tract diseases (with RR = 1.016 at lag 06, 95% CI: 1.015–1.017, per 10 μg/m3 PM10 increase, RR = 1.073 at lag 06, 95% CI: 1.070–1.076, per 10 ppb O3 increase, and RR = 1.046 at lag 06, 95% CI: 1.040–1.051, per 10 ppb NO2 increase). Multi-pollutants modeling analysis identified O3 as a relatively independent risk factor and PM10-NO2 pollutants models as promising two-pollutants models. Overall, these results demonstrate the adverse effects of all three air pollutants on respiratory morbidity and call for air pollution reduction and control.


  • Open Access

    A real-world study of effectiveness of intravitreal bevacizumab and ranibizumab injection for treating retinal diseases in Thailand

    by Suthasinee Kumluang, Lily Ingsrisawang, Sermsiri Sangroongruangsri, Usa Chaikledkaew, Tanapat Ratanapakorn, Paisan Ruamviboonsuk, Wongsiri Taweebanjongsin, Janejit Choovuthayakorn, Somanus Thoongsuwan, Prut Hanutsaha, et al.

    To evaluate the effectiveness of intravitreal bevacizumab (IVB) and intravitreal ranibizumab (IVR) in actual practice for treating patients with retinal diseases in Thailand. A prospective, multi-centre, observational study was conducted among eight hospitals in their ophthalmology outpatient departments. Participants consisted of patients who had previously not received any IVB or IVR treatment between 2013 and 2014. The primary outcome measurement was the change in best-corrected visual acuity (BCVA) at the end of the follow-up period compared to baseline. There were 1629 treatment-naïve patients for the pro re nata (PRN) treatment pattern and 226 treatment-naive patients for the three-injections (3Inj) treatment pattern. BCVA improvements were found in 35% of the PRN group and 47% of the 3Inj group; however, it was not clinically meaningful between the IVB and IVR groups (P-value = 0.568 for PRN, P-value = 0.103 for 3Inj). A multivariable logistic regression (using the propensity score) showed that positive factors associated with vision improvement for the PRN pattern were the number of drug injections, having retinal vein occlusion, and under 60 years of age, while good BCVA at baseline was a negative predictive factor. For the 3Inj pattern, under 60 years of age and baseline BCVA were statistically significant predictors. Nonetheless, diabetes mellitus (DM) without other comorbidities was a statistically significant predictor of low response to vision improvement compared to DM with other comorbidities. This study was the first observational, prospective study to evaluate the real-life effectiveness of IVB and IVR in Thailand. The majority of participants who used IVB or IVR showed improvements in BCVA after treatment. Further evaluation such as long-term follow-ups and subsequent comparison of effectiveness between IVB and IVR should be investigated due to the limited sample of IVR patients. Thai Clinical Trial Registry TCTR20141002001 . Registered 02 October 2014 (retrospectively registered).


  • Contribution of Different Antiretroviral Regimens Containing Zidovudine, Lamivudine and Ritonavir-Boosted Lopinavir on HIV Viral Load Reduction during Pregnancy

    by Patumrat Sripan, Sophie Le Coeur, Lily Ingsrisawang, Tim R Cressey, Naïm Bouazza, Frantz Foissac, Nicole Ngo-Giang-Huong, Patrinee Traisathit, Ussanee Srirompotong, Orada Patamasingh Na Ayudhaya, et al.

    Background: Antiretroviral (ARV) regimens used for the prevention of mother-to-child transmission of HIV have evolved over time. We evaluated the contribution of different ARV regimens on the reduction of the plasma HIV RNA viral load (VL) during pregnancy. Methods: A total of 1,833 VL measurements from ARV-naive pregnant women participating in perinatal prevention trials in Thailand were included. Women received either zidovudine (ZDV) monotherapy, ZDV plus lopinavir/ritonavir (LPV/r), or ZDV plus lamivudine (3TC) plus LPV/r. VL time-course during pregnancy was described as a function of pretreatment VL and treatment duration using an E max non-linear mixed-effect model. VL reduction and median time to achieve a VL≤50 copies/ml were estimated for each regimen. Results: Among 745 women, 279 (37%), 145 (20%) and 321 (43%) received ZDV monotherapy, ZDV+LPV/r and ZDV+3TC+LPV/r, respectively. The predicted VL reduction from baseline to delivery after a median of 10 weeks of treatment were 0.5, 2.7 and 2.9 log 10 copies/ml with ZDV monotherapy, ZDV+LPV/r and ZDV+3TC+LPV/r, respectively. At delivery, 1%, 57% and 63% of women receiving ZDV monotherapy, ZDV+LPV/r or ZDV+3TC+LPV/r had a VL≤50 copies/ml. The addition of 3TC to ZDV+LPV/r reduced the time to achieve a VL≤50 copies/ml and the higher the pretreatment VL, the larger the effect 3TC had on reducing the time to VL≤50 copies/ml. Conclusions: The addition of 3TC to ZDV+LPV/r was associated with a slight further VL reduction but the time to reach a VL≤50 copies/ml was shorter. This beneficial effect of 3TC is crucial for prevention of mother-to-child transmission in women who receive ARVs late and with high pretreatment VL.