Not all MDAs should be created equal – determinants of MDA impact and designing MDAs towards malaria elimination

Malaria remains at the forefront of scientific research and global political and funding agendas. Previous malaria models of mass-interventions have consistently oversimplified how mass interventions are implemented. We present an individual based, spatially explicit model of malaria transmission that includes all the programmatic implementation details of mass drug administration (MDA) campaigns. We uncover how the impact of MDA campaigns is determined by the interaction between implementation logistics, patterns of human mobility and how transmission risk is distributed over space. This translates into a higher likelihood of malaria elimination for areas with true prevalence under 3% with a faster implementation, in highly mobile populations. If populations are more static, deploying less interventions teams would be cost optimal and predicted to be as impactful. We conclude that mass drug interventions can be an invaluable tool towards malaria elimination in the right context, specifically when paired with effective vector control.


Introduction 26
In Southeast Asia, and particularly the Greater Mekong Sub-region (GMS), Plasmodium 27 falciparum transmission has decreased substantially over the last two decades (1,2), setting the 28 stage for pre-elimination scenarios, with all GMS countries committing to ambitious elimination parasites) and logistical constraints -Table 1 -modulate the expected outcome of MDA campaigns. Figure 2 illustrates the sensitivity of the predicted proportional decrease in prevalence 93 over 5 years to each parameter. Clearly, the number of MDA rounds and the initial mean 94 prevalence across all villages are critical covariates when predicting MDA outcome. Also quite 95 important seem to be the distributions characterizing how malaria risk is distributed over space.

96
Artemisinin resistance spread is quite sensitive to the same covariates and additionally to the

99
We found that there is an intricate relationship between the optimal timing of MDA campaign 100 start, its implementation logistics, and malaria seasonality patterns. Deploying a higher number of 101 MDA teams will yield a higher likelihood of reaching malaria elimination within 2 years, only 102 when population mobility is high - Figure 3. Using a smaller number of intervention teams is 103 predicted to be advantageous in more static population, especially when there is only one annual less pronounced in populations with high mobility, due to a dilution effect described below.

169
A slower deployment of MDA campaigns is then generally preferable in low mobility

204
When there is only one transmission peak during the year, a slower MDA implementation 205 seems to be preferable (Figure 3), especially if the start of the MDA is set to start one month 206 prior to the peak in vectorial capacity (not to be mistaken with the season malaria incidence 207 peak which occurs later) instead of starting at the beginning of the year. It seems delaying 208 the start of MDA campaigns improves the likelihood of malaria elimination compared to a 209 start at the beginning of the year when a lower number of MDA teams is used (Figure 3 -210 supplement 1). This is mostly due to how sensitive near instantaneous MDA campaigns 211 are to the timing of the seasonal peak. In a single annual peak setting, where the incidence 212 peak is at day 140, a near instantaneous MDA would end (all 3 rounds) 70 (for no delay) or 10 (with delay) days prior to the peak. Given the general cosine function simulated here,

214
it seems a longer implementation lasting the whole duration of the high transmission season whole of the first annual peak) seems preferable.

217
 Dilution. We uncovered an interesting trade-off between population mobility, transmission 218 heterogeneity and number of MDA teams that results in unexpectedly high predictions for 219 intervention impact in highly mobile populations. This is due to a diluting effect, rooted in

237
Interestingly, this effect is more pronounced for populations with the lowest mobility for 238 all considered spatial heterogeneity distributions. This is due to the dilution effect

299
In fact, the most skewed distribution offers better elimination prospects for higher negligible vectorial capacity, transmission is sustained by a few high transmission foci. If 303 those are targeted efficiently, you can expect to disrupt transmission more than in settings 304 where the distribution of mosquitoes over space is more homogeneous.

306
We have refrained from doing a cost-effectiveness analysis, since we do not have good information 307 on most of the unit costs, which are currently being assessed in different settings, and are likely to 308 be quite variable across countries in the GMS. Any recommendation and cost-effectiveness 309 analysis would have to be tailored to each specific country/area. We also have not extensively treatments, but if that is the case, then it would manifest itself at the time of the prevalence survey, when blood would have to be drawn. In practice, we concede that we may be unable to deploy MDA in whole villages due to these constraints, but in the absence of data we refrain from making 322 any assumptions.

323
Here, we present a theoretical exploration of the potential impact of MDA strategies in different 324 settings of the GMS, with special emphasis on the sensitivity of the predicted impact to logistical 325 constraints, and transmission or population topologies. The ranges of parameters and distributions 326 explored are meant to represent the current malaria situation in the GMS but need to be adjusted 327 for application to specific areas/countries. In conclusion, we propose that mass drug interventions 328 can be an invaluable tool towards malaria elimination in the right context. The model presented 329 here predicts that an MDA's success likelihood is bounded by the initial malaria prevalence and 330 we elucidate how those chances can be improved through tailoring of implementation logistics.

331
Although MDA is being revisited by the global community, very little attention has been paid to 332 implementation logistics, and there seems to be no protocol adjustment across settings with       The developed malaria micro-simulation platform takes inputs from two CSV-formatted data files and a 482 JSON-formatted configuration file. The two data files provide the model with a list of villages and a list of 483 humans respectively. Each row of these two data files describes the properties (as listed in the previous 484 section) of either a village or a human agent. Model-wide and process-specific, as opposed to agent-specific, 485 parameters are given by the configuration file. Most parameters specified in Table 2 of the main text are 486 associated with processes and functions (rather than individual agents), and therefore are given by the 487 configuration file. In this section, we use the numbers in square brackets to refer to the associated parameter 488 number whose description and value can be found in Table 2. 489 The initialisation process starts by processing the configuration file to get information such as the location 490 of the other input files. Then a list of village agents and a list of human agents are created according to 491 information given in the data files. Human agents are either randomly assigned a home village from all 492 available villages or assigned to villages in the same order as they appear in the input data file. Once all 493 agents have been created, the software initialises the parameterised-functions of the model with the 494 information given in the rest of the configuration file. 495

Human Properties
Note that whereas the configuration file and the village data file are mandatory, the data file for humans is 498 optional to the model software. Given that the population of each village is known from the village data 499 file, in the absence of a human data file, a set of human agents (N [2]) may be generated by the software 500 according to each village's population size. We now give details of the processes that are part of the 501 initialisation process when properties of human agents are not specified by input. 502

Gender and Age 503
Age and gender information of human agents are generated according to parameterised distributions. The 504 probability of a human agent being male is ml [3]. The age of a human is sampled from a discrete 505 distribution specified by a vector of size maxage [4]. This vector is given by a csv file with maxage [4] 506

integers. 507
Prevalence 508 The probability of a human agent being infectious is previ [6], and the probability of that infection being 509 resistant to artemisinin is resit [7]. 510

ITN Usage 511
Each human agent is assigned with an ITN with a probability of netuse The simulation protocols detailed in this section illustrate the flow of events and processes taking place 537 each day. These dynamic processes are reduced to daily probabilities of occurrence, with events occurring 538 if a randomly drawn, uniformly distributed number between 0 and 1, is lower than the likelihood of 539 something happening on a given day. Thus, for each individual and for every simulated daily time step, a 540 random uniformly distributed number is drawn for each possible transition process (e.g. whether the 541 individual dies, is infected, is treated, etc), with that specific transition taking place if the number drawn is 542 lower than the respective daily probability of occurrence. 543

544
Birth and death 545 The probability for the death of a human agent is given by 546 where agen denotes the age property of the human agent ∈ . When a death event happens, a new agent 548 is generated as a replacement. The new agent is placed in the same village where the death took place to 549 keep population size a constant. 550

Short-term movement 551
Every human agent can display short-term mobility patterns, characterised by overnight stays in villages 552 other than their home for a mean period of ovstay [10] days. For every agent who is currently located at 553 their home village, the daily probability of such short-term movement is given by the factor Mobility in 554 Table 1. Thus, the number of human agents embarking on short-term movement on a given day is 555 The destination of each agent's movement varies and is determined using the mobility network 557 constructed during initialisation. Let the flow of short-term movement between village , ∈ be 558 ) ) 559 where crit [11] denotes the critical distance below which overnight stays at a village other than home are 560 made very unlikely. The probability of a human agent to move from ℎ to village is 561 Clinical outcome 563 The likelihood of clinical symptoms brought on by a single infection is given by 564

565
where moi, cml and lvl denote the multiplicity of infection, cumulative exposure to malaria and immunity 566 level properties of the human agent respectively. 567

Treatment 568
The probability for a human agent to receive a full course of ACT treatment is dependent on 569 symptomatology as well as the presence of a local malaria post. In a village with a malaria post, a human 570 agent with clinical symptom would receive treatment with probability 571 Parasite killing rates depend on the person's transmission status ( ), with parasite clearance in not yet 593 infectious people generally slower than that in individuals carrying gametocytes. Clearance of parasites 594 with drug resistance phenotype ℎ by drug then follows 595 denotes a binomial 597

distribution. 598
Recovery 599 Each infection in a human agent's infection list has a daily probability of being naturally cleared given by 600

1/delta [25]. 601
Immunity 602 One level of clinical immunity is gained by a human agent every time his infection list is emptied. Immunity 603 loss starts imm_min [26] days after one level of immunity is gained. Immunity is lost at a rate of 1/alpha 604 In order to coordinate MDA teams to visit all villages without overlapping and repetition, a complete graph 651 connecting all villages to each other is first constructed. The weight of the edge between village and 652 is given by ( , ). This complete graph is then reduced to its minimum spanning tree form, denoted 653 ( ), which we use to represent the road network connecting all villages. 654 Given that the MDA campaign includes teams (Table 1), starting locations are randomly selected from 655 the nodes/villages of ( ). Then, breadth-first-search algorithms are started from each of the 656 starting locations. These search algorithms run simultaneously in coordinated rounds. Each round, an 657 algorithm is to add a village to its path. A village is added to the path of the algorithm which reaches it first. 658 When an algorithm reaches a village that has been added/visited to the path of another algorithm, its                 Figure 5, except that vector control is extended to last 5 years.